From e5ae08a364c459548e7c5a7d28e46421fe18db33 Mon Sep 17 00:00:00 2001 From: Quarto GHA Workflow Runner Date: Fri, 8 Aug 2025 01:27:51 +0000 Subject: [PATCH] Built site for gh-pages --- .nojekyll | 2 +- docs/api/prompt_strategies.chat_template.html | 10 +- docs/api/utils.collators.batching.html | 3 +- .../utils.ctx_managers.sequence_parallel.html | 3 +- docs/config-reference.html | 2170 +++++++++-------- docs/dataset-formats/conversation.html | 3 +- search.json | 18 +- sitemap.xml | 396 +-- 8 files changed, 1311 insertions(+), 1294 deletions(-) diff --git a/.nojekyll b/.nojekyll index 3d1d27dc9..fc07244db 100644 --- a/.nojekyll +++ b/.nojekyll @@ -1 +1 @@ -d0e15477 \ No newline at end of file +a51154a6 \ No newline at end of file diff --git a/docs/api/prompt_strategies.chat_template.html b/docs/api/prompt_strategies.chat_template.html index 93260bccc..e967caaec 100644 --- a/docs/api/prompt_strategies.chat_template.html +++ b/docs/api/prompt_strategies.chat_template.html @@ -560,10 +560,12 @@ gtag('config', 'G-9KYCVJBNMQ', { 'anonymize_ip': true}); field_messages='messages', field_system='system', field_tools='tools', - roles=None, - chat_template_kwargs=None, - drop_system_message=False, -) + field_thinking='reasoning_content', + roles=None, + template_thinking_key='reasoning_content', + chat_template_kwargs=None, + drop_system_message=False, +)

Prompter for HF chat templates

Methods

diff --git a/docs/api/utils.collators.batching.html b/docs/api/utils.collators.batching.html index ac5cf30a4..90a049747 100644 --- a/docs/api/utils.collators.batching.html +++ b/docs/api/utils.collators.batching.html @@ -653,7 +653,8 @@ gtag('config', 'G-9KYCVJBNMQ', { 'anonymize_ip': true}); label_pad_token_id=-100, position_pad_token_id=0, return_tensors='pt', -) + squash_position_ids=False, +)

Collator for multipack specific to the using the BatchSampler

diff --git a/docs/api/utils.ctx_managers.sequence_parallel.html b/docs/api/utils.ctx_managers.sequence_parallel.html index 8466744c1..7f3463354 100644 --- a/docs/api/utils.ctx_managers.sequence_parallel.html +++ b/docs/api/utils.ctx_managers.sequence_parallel.html @@ -710,7 +710,8 @@ from the full gradient tensor.

ring_attn_func, heads_k_stride, gather_outputs, -) + device_mesh=None, +)

Context manager for sequence parallelism operations.

This class provides a context that will automatically apply sequence parallelism during model forward passes using a pre-forward hook, and gather outputs from diff --git a/docs/config-reference.html b/docs/config-reference.html index 7abc60bd6..2f7953e9e 100644 --- a/docs/config-reference.html +++ b/docs/config-reference.html @@ -740,1090 +740,1102 @@ gtag('config', 'G-9KYCVJBNMQ', { 'anonymize_ip': true}); # Key containing the tools (default: "tools"). Must be a list[dict] and follow [JSON # schema](https://json-schema.org/learn/getting-started-step-by-step). field_tools: str | None - - message_field_role: str | None - - message_field_content: str | None - # Mapping of properties from the input dataset to the chat template. (default: - # message_property_mappings={'role':'role', 'content':'content'}) If a property exists - # in the template but not in this mapping, the system will attempt to load it directly - # from the message using the property name as the key. Example: In the mapping below, - # 'from' is loaded from input dataset and used as 'role', while 'value' is loaded and - # used as 'content' in the chat template. - message_property_mappings: dict[str, str] | None - # The key in the message turn that indicates via boolean whether tokens of a turn - # should be considered for training. Useful to selectively train on certain turns - # besides the `roles_to_train`. - message_field_training: str | None - # The key in the message turn that contains the training details. Useful to - # selectively train on certain tokens in a turn. The value of the key is a List[Dict] - # containing `begin_offset` (start character index in content), `end_offset` (end - # character index in content), and `train` (boolean whether to train). - message_field_training_detail: str | None - # (for Qwen3 template only) Whether to split the assistant content based on a - # reasoning trace inside delimited tags - split_thinking: bool | None - logprobs_field: str | None - temperature: float | None - # Roles to train on. The tokens from these roles will be considered for the loss. - roles_to_train: list[str] | None - # Which EOS tokens to train on in the conversation. Possible values are: all: train on - # all EOS tokens, turn (default): train on the EOS token at the end of each trainable - # turn, last: train on the last EOS token in the conversation - train_on_eos: Literal['all', 'turn', 'last'] | None - # Roles mapping in the messages. The format is {target_role: [source_roles]}. All - # source roles will be mapped to the target role. The default is: user: ["human", - # "user"], assistant: ["gpt", "assistant"], system: ["system"], tool: ["tool"] - roles: dict[str, list[str]] | None - # Whether to drop the system turn from the dataset. Only works with chat_template. - # This does not drop the default system message from chat_template if it exists. If - # you wish to, we recommend using a custom jinja template with the default system - # message removed or adding a system turn with empty content. - drop_system_message: bool | None - # Trust remote code for untrusted source - trust_remote_code: bool | None = False - # The specific revision of the dataset to use when loading from the Hugging Face Hub. - # This can be a commit hash, tag, or branch name. If not specified, the latest version - # will be used. This parameter is ignored for local datasets. - revision: str | None - - # For DPODataset: - path: str | None - split: str | None - type: UserDefinedDPOType | str | None - # For UserDefinedDPOType: - field_system: str | None - field_prompt: str | None - field_chosen: str | None - field_rejected: str | None - prompt_format: str | None - chosen_format: str | None - rejected_format: str | None - data_files: list[str] | None - revision: str | None - field_messages: str | None - - # For KTODataset: - path: str | None - split: str | None - type: UserDefinedKTOType | str | None - # For UserDefinedKTOType: - field_system: str | None - field_prompt: str | None - field_completion: str | None - field_label: bool | None - prompt_format: str | None - completion_format: str | None - data_files: list[str] | None - trust_remote_code: bool | None = False - revision: str | None - - # For StepwiseSupervisedDataset: - path: str | None - split: str | None - data_files: list[str] | None - revision: str | None - step_separator: str | None - max_completion_length: int | None - train_on_last_step_only: bool | None - -# A list of one or more datasets to eval the model with. You can use either -# test_datasets, or val_set_size, but not both. -test_datasets: Annotated[list[SFTDataset | DPODataset | KTODataset | StepwiseSupervisedDataset], MinLen(1)] | None - # For SFTDataset: - # HuggingFace dataset repo | s3:// | gs:// | path to local file or directory - path: str | None - # name of dataset split to load from - split: str | None - # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection] - type: str | UserDefinedPrompterType | None - # For UserDefinedPrompterType: - # Custom user instruction prompt - system_prompt: str | None - # Use {system} as key to be replaced - system_format: str | None - field_system: str | None - field_instruction: str | None - field_input: str | None - field_output: str | None - - # Customizable to be single line or multi-line. Use {instruction}/{input} as key to - # be replaced. 'format' can include {input} - format: str | None - # 'no_input_format' cannot include {input} - no_input_format: str | None - input_transform: str | None - # split dataset into N pieces (use with shards_idx) - shards: int | None - # the index of sharded dataset to use - shards_idx: int | None - # process dataset in N sequential chunks for memory efficiency (exclusive with - # `shards`) - preprocess_shards: int | None - conversation: str | None - - # The name of the chat template to use for training, following values are supported: - # tokenizer_default: Uses the chat template that is available in the - # tokenizer_config.json. If the chat template is not available in the tokenizer, it - # will raise an error. This is the default. - # alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates - # are available in the axolotl codebase at src/axolotl/utils/chat_templates.py. - # tokenizer_default_fallback_*: where * is the name of the chat template to fallback - # to if the tokenizer does not have a chat template else default to tokenizer. E.g. - # tokenizer_default_fallback_chatml. jinja: Uses a custom jinja template for the chat - # template. The custom jinja template should be provided in the chat_template_jinja - # field. - chat_template: ChatTemplate | str | None - # Custom jinja chat template or path to jinja file. Used only if `chat_template: - # jinja` or empty. - chat_template_jinja: str | None - # path to source data files - data_files: str | list[str] | None - input_format: str | None - # name of dataset configuration to load - name: str | None - # defines the datatype when path is a file - ds_type: str | None - # For `completion` datasets only, uses the provided field instead of `text` column - field: str | None - field_human: str | None - field_model: str | None - # Key containing the messages (default: "messages") - field_messages: str | None - # Key containing the tools (default: "tools"). Must be a list[dict] and follow [JSON - # schema](https://json-schema.org/learn/getting-started-step-by-step). - field_tools: str | None - - message_field_role: str | None - - message_field_content: str | None - # Mapping of properties from the input dataset to the chat template. (default: - # message_property_mappings={'role':'role', 'content':'content'}) If a property exists - # in the template but not in this mapping, the system will attempt to load it directly - # from the message using the property name as the key. Example: In the mapping below, - # 'from' is loaded from input dataset and used as 'role', while 'value' is loaded and - # used as 'content' in the chat template. - message_property_mappings: dict[str, str] | None - # The key in the message turn that indicates via boolean whether tokens of a turn - # should be considered for training. Useful to selectively train on certain turns - # besides the `roles_to_train`. - message_field_training: str | None - # The key in the message turn that contains the training details. Useful to - # selectively train on certain tokens in a turn. The value of the key is a List[Dict] - # containing `begin_offset` (start character index in content), `end_offset` (end - # character index in content), and `train` (boolean whether to train). - message_field_training_detail: str | None - # (for Qwen3 template only) Whether to split the assistant content based on a - # reasoning trace inside delimited tags - split_thinking: bool | None - logprobs_field: str | None - temperature: float | None - # Roles to train on. The tokens from these roles will be considered for the loss. - roles_to_train: list[str] | None - # Which EOS tokens to train on in the conversation. Possible values are: all: train on - # all EOS tokens, turn (default): train on the EOS token at the end of each trainable - # turn, last: train on the last EOS token in the conversation - train_on_eos: Literal['all', 'turn', 'last'] | None - # Roles mapping in the messages. The format is {target_role: [source_roles]}. All - # source roles will be mapped to the target role. The default is: user: ["human", - # "user"], assistant: ["gpt", "assistant"], system: ["system"], tool: ["tool"] - roles: dict[str, list[str]] | None - # Whether to drop the system turn from the dataset. Only works with chat_template. - # This does not drop the default system message from chat_template if it exists. If - # you wish to, we recommend using a custom jinja template with the default system - # message removed or adding a system turn with empty content. - drop_system_message: bool | None - # Trust remote code for untrusted source - trust_remote_code: bool | None = False - # The specific revision of the dataset to use when loading from the Hugging Face Hub. - # This can be a commit hash, tag, or branch name. If not specified, the latest version - # will be used. This parameter is ignored for local datasets. - revision: str | None - - # For DPODataset: - path: str | None - split: str | None - type: UserDefinedDPOType | str | None - # For UserDefinedDPOType: - field_system: str | None - field_prompt: str | None - field_chosen: str | None - field_rejected: str | None - prompt_format: str | None - chosen_format: str | None - rejected_format: str | None - data_files: list[str] | None - revision: str | None - field_messages: str | None - - # For KTODataset: - path: str | None - split: str | None - type: UserDefinedKTOType | str | None - # For UserDefinedKTOType: - field_system: str | None - field_prompt: str | None - field_completion: str | None - field_label: bool | None - prompt_format: str | None - completion_format: str | None - data_files: list[str] | None - trust_remote_code: bool | None = False - revision: str | None - - # For StepwiseSupervisedDataset: - path: str | None - split: str | None - data_files: list[str] | None - revision: str | None - step_separator: str | None - max_completion_length: int | None - train_on_last_step_only: bool | None - -# If false, the datasets will not be shuffled and will keep their original order in -# `datasets`. The same applies to the `test_datasets` option and the -# `pretraining_dataset` option. Default is true. -shuffle_merged_datasets: bool | None = True -# If true, each dataset in `datasets` will be shuffled before merging. This allows -# curriculum learning strategies to be applied at the dataset level. Default is false. -shuffle_before_merging_datasets: bool | None = False -# Axolotl attempts to save the dataset as an arrow after packing the data together so -# subsequent training attempts load faster, relative path -dataset_prepared_path: str | None -# Num shards for whole dataset -dataset_shard_num: int | None -# Index of shard to use for whole dataset -dataset_shard_idx: int | None -skip_prepare_dataset: bool | None = False -# Number of shards to save the prepared dataset -num_dataset_shards_to_save: int | None - -# Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize -pretraining_dataset: Annotated[list[PretrainingDataset | SFTDataset], MinLen(1)] | None - # For PretrainingDataset: - name: str | None - path: str | None - split: str | None = train - text_column: str | None = text - type: str | None = pretrain - trust_remote_code: bool | None = False - data_files: str | None - skip: int | None - - # For SFTDataset: - # HuggingFace dataset repo | s3:// | gs:// | path to local file or directory - path: str | None - # name of dataset split to load from - split: str | None - # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection] - type: str | UserDefinedPrompterType | None - # For UserDefinedPrompterType: - # Custom user instruction prompt - system_prompt: str | None - # Use {system} as key to be replaced - system_format: str | None - field_system: str | None - field_instruction: str | None - field_input: str | None - field_output: str | None - - # Customizable to be single line or multi-line. Use {instruction}/{input} as key to - # be replaced. 'format' can include {input} - format: str | None - # 'no_input_format' cannot include {input} - no_input_format: str | None - input_transform: str | None - # split dataset into N pieces (use with shards_idx) - shards: int | None - # the index of sharded dataset to use - shards_idx: int | None - # process dataset in N sequential chunks for memory efficiency (exclusive with - # `shards`) - preprocess_shards: int | None - conversation: str | None - - # The name of the chat template to use for training, following values are supported: - # tokenizer_default: Uses the chat template that is available in the - # tokenizer_config.json. If the chat template is not available in the tokenizer, it - # will raise an error. This is the default. - # alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates - # are available in the axolotl codebase at src/axolotl/utils/chat_templates.py. - # tokenizer_default_fallback_*: where * is the name of the chat template to fallback - # to if the tokenizer does not have a chat template else default to tokenizer. E.g. - # tokenizer_default_fallback_chatml. jinja: Uses a custom jinja template for the chat - # template. The custom jinja template should be provided in the chat_template_jinja - # field. - chat_template: ChatTemplate | str | None - # Custom jinja chat template or path to jinja file. Used only if `chat_template: - # jinja` or empty. - chat_template_jinja: str | None - # path to source data files - data_files: str | list[str] | None - input_format: str | None - # name of dataset configuration to load - name: str | None - # defines the datatype when path is a file - ds_type: str | None - # For `completion` datasets only, uses the provided field instead of `text` column - field: str | None - field_human: str | None - field_model: str | None - # Key containing the messages (default: "messages") - field_messages: str | None - # Key containing the tools (default: "tools"). Must be a list[dict] and follow [JSON - # schema](https://json-schema.org/learn/getting-started-step-by-step). - field_tools: str | None - - message_field_role: str | None - - message_field_content: str | None - # Mapping of properties from the input dataset to the chat template. (default: - # message_property_mappings={'role':'role', 'content':'content'}) If a property exists - # in the template but not in this mapping, the system will attempt to load it directly - # from the message using the property name as the key. Example: In the mapping below, - # 'from' is loaded from input dataset and used as 'role', while 'value' is loaded and - # used as 'content' in the chat template. - message_property_mappings: dict[str, str] | None - # The key in the message turn that indicates via boolean whether tokens of a turn - # should be considered for training. Useful to selectively train on certain turns - # besides the `roles_to_train`. - message_field_training: str | None - # The key in the message turn that contains the training details. Useful to - # selectively train on certain tokens in a turn. The value of the key is a List[Dict] - # containing `begin_offset` (start character index in content), `end_offset` (end - # character index in content), and `train` (boolean whether to train). - message_field_training_detail: str | None - # (for Qwen3 template only) Whether to split the assistant content based on a - # reasoning trace inside delimited tags - split_thinking: bool | None - logprobs_field: str | None - temperature: float | None - # Roles to train on. The tokens from these roles will be considered for the loss. - roles_to_train: list[str] | None - # Which EOS tokens to train on in the conversation. Possible values are: all: train on - # all EOS tokens, turn (default): train on the EOS token at the end of each trainable - # turn, last: train on the last EOS token in the conversation - train_on_eos: Literal['all', 'turn', 'last'] | None - # Roles mapping in the messages. The format is {target_role: [source_roles]}. All - # source roles will be mapped to the target role. The default is: user: ["human", - # "user"], assistant: ["gpt", "assistant"], system: ["system"], tool: ["tool"] - roles: dict[str, list[str]] | None - # Whether to drop the system turn from the dataset. Only works with chat_template. - # This does not drop the default system message from chat_template if it exists. If - # you wish to, we recommend using a custom jinja template with the default system - # message removed or adding a system turn with empty content. - drop_system_message: bool | None - # Trust remote code for untrusted source - trust_remote_code: bool | None = False - # The specific revision of the dataset to use when loading from the Hugging Face Hub. - # This can be a commit hash, tag, or branch name. If not specified, the latest version - # will be used. This parameter is ignored for local datasets. - revision: str | None - -# The maximum number of processes to use while preprocessing your input dataset. This -# defaults to `os.cpu_count()` if not set. For Runpod VMs, it will default to number of -# vCPUs via RUNPOD_CPU_COUNT. -dataset_processes: int | None -# Deduplicates datasets and test_datasets with identical entries -dataset_exact_deduplication: bool | None -# Keep dataset in memory while preprocessing. Only needed if cached dataset is taking -# too much storage -dataset_keep_in_memory: bool | None -dataloader_pin_memory: bool | None -dataloader_num_workers: int | None -dataloader_prefetch_factor: int | None -dataloader_drop_last: bool | None - -accelerator_config: dict[str, Any] | None - -remove_unused_columns: bool | None - -# Push prepared dataset to hub - repo_org/repo_name -push_dataset_to_hub: str | None -# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private -# datasets. Required to be true when used in combination with `push_dataset_to_hub` -hf_use_auth_token: bool | None - -device: Any | None -# Passed through to transformers when loading the model when launched without -# accelerate. Use `sequential` when training w/ model parallelism to limit memory -device_map: Any | None -world_size: int | None -# Don't mess with this, it's here for accelerate and torchrun -local_rank: int | None -ddp: bool | None - -# Seed for reproducibility -seed: int | None -# Advanced DDP Arguments - timeout -ddp_timeout: int | None -# Advanced DDP Arguments - bucket cap in MB -ddp_bucket_cap_mb: int | None -# Advanced DDP Arguments - broadcast buffers -ddp_broadcast_buffers: bool | None -ddp_find_unused_parameters: bool | None - -# Approximate number of predictions sent to wandb depending on batch size. Enabled above -# 0. Default is 0 -eval_table_size: int | None -# Total number of tokens generated for predictions sent to wandb. Default is 128 -eval_max_new_tokens: int | None -# Whether to run causal language model evaluation for metrics in -# `eval_causal_lm_metrics` -do_causal_lm_eval: bool | None -# HF evaluate metrics used during evaluation. Default is ['sacrebleu', 'comet', 'ter', -# 'chrf', 'perplexity'] -eval_causal_lm_metrics: list[str] | None -do_bench_eval: bool | None -bench_dataset: str | None -bench_split: str | None -metric_for_best_model: str | None -greater_is_better: bool | None - -# High loss value, indicating the learning has broken down (a good estimate is ~2 times -# the loss at the start of training) -loss_watchdog_threshold: float | None -# Number of high-loss steps in a row before the trainer aborts (default: 3) -loss_watchdog_patience: int | None - -# Run garbage collection every `gc_steps` steps. -1 will run on epoch end and before -# evaluations. Default is 0 (disabled). -gc_steps: int | None - -# Use CUDA bf16. bool or 'full' for `bf16_full_eval`, or 'auto' for automatic detection. -# require >=ampere -bf16: Literal['auto'] | bool | None = auto -# Use CUDA fp16 -fp16: bool | None -# Enable FP8 mixed precision training using TorchAO. Best used in combination with -# torch.compile. -fp8: bool | None -# Enable FSDP float8 all-gather optimization for FP8 training. Can improve training -# speed by 10-15% when FSDP is enabled. -fp8_enable_fsdp_float8_all_gather: bool | None -# No AMP (automatic mixed precision) - require >=ampere -bfloat16: bool | None -# No AMP (automatic mixed precision) -float16: bool | None -# Use CUDA tf32 - require >=ampere -tf32: bool | None -float32: bool | None - -# Whether to use gradient checkpointing. Available options are: true, false, 'offload', -# 'offload_disk'. -# https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing -gradient_checkpointing: Literal['offload', 'offload_disk'] | bool | None = False -# Additional kwargs to pass to the trainer for gradient checkpointing -gradient_checkpointing_kwargs: dict[str, Any] | None -# Whether to offload activations. Available options are: true, false, 'legacy', 'disk'. -activation_offloading: Literal['legacy', 'disk'] | bool | None = False - -unfrozen_parameters: list[str] | None - -# The maximum length of an input to train with, this should typically be less than 2048 -# as most models have a token/context limit of 2048 -sequence_len: int = 512 -# The maximum length of an input for evaluation. If not specified, defaults to -# sequence_len -eval_sequence_len: int | None -min_sample_len: int | None -# maximum prompt length for RL training -max_prompt_len: int = 512 -# Use efficient multi-packing with block diagonal attention and per sequence -# position_ids. Recommend set to 'true' -sample_packing: bool | None -# The number of samples packed at a time. Increasing the following values helps with -# packing, but usually only slightly (<%1.) -sample_packing_group_size: int | None = 100000 -# The number of samples which can be packed into one sequence. Increase if using a large -# sequence_len with many short samples. -sample_packing_bin_size: int | None = 200 -# Whether to pack samples sequentially -sample_packing_sequentially: bool | None -# The multiprocessing start method to use for packing. Should be 'fork', 'spawn' or -# 'forkserver' -sample_packing_mp_start_method: str | None -# Set to 'false' if getting errors during eval with sample_packing on -eval_sample_packing: bool | None -# Pad inputs so each step uses constant sized buffers. This will reduce memory -# fragmentation and may prevent OOMs, by re-using memory more efficiently. Defaults to -# True if `sample_packing` enabled -pad_to_sequence_len: bool | None -# Whether to use sequential sampling for curriculum learning -curriculum_sampling: bool | None -multipack_real_batches: bool | None -# whether to concatenate samples during pretraining -pretraining_sample_concatenation: bool | None - -# Use batch flattening for speedups when not using sample_packing -batch_flattening: Literal['auto'] | bool | None - -use_pose: bool | None -pose_split_on_token_ids: list[int] | None -pose_max_context_len: int | None -pose_num_chunks: int | None - -pretrain_multipack_buffer_size: int | None = 10000 -# whether to prevent cross attention for packed sequences during pretraining -pretrain_multipack_attn: bool | None = True + # Key containing the reasoning trace (default: "reasoning_content"). + field_thinking: str | None + # The key the chat template expects that indicates the reasoning trace. + template_thinking_key: str | None + + message_field_role: str | None + + message_field_content: str | None + # Mapping of properties from the input dataset to the chat template. (default: + # message_property_mappings={'role':'role', 'content':'content'}) If a property exists + # in the template but not in this mapping, the system will attempt to load it directly + # from the message using the property name as the key. Example: In the mapping below, + # 'from' is loaded from input dataset and used as 'role', while 'value' is loaded and + # used as 'content' in the chat template. + message_property_mappings: dict[str, str] | None + # The key in the message turn that indicates via boolean whether tokens of a turn + # should be considered for training. Useful to selectively train on certain turns + # besides the `roles_to_train`. + message_field_training: str | None + # The key in the message turn that contains the training details. Useful to + # selectively train on certain tokens in a turn. The value of the key is a List[Dict] + # containing `begin_offset` (start character index in content), `end_offset` (end + # character index in content), and `train` (boolean whether to train). + message_field_training_detail: str | None + # (for Qwen3 template only) Whether to split the assistant content based on a + # reasoning trace inside delimited tags + split_thinking: bool | None + logprobs_field: str | None + temperature: float | None + # Roles to train on. The tokens from these roles will be considered for the loss. + roles_to_train: list[str] | None + # Which EOS tokens to train on in the conversation. Possible values are: all: train on + # all EOS tokens, turn (default): train on the EOS token at the end of each trainable + # turn, last: train on the last EOS token in the conversation + train_on_eos: Literal['all', 'turn', 'last'] | None + # Roles mapping in the messages. The format is {target_role: [source_roles]}. All + # source roles will be mapped to the target role. The default is: user: ["human", + # "user"], assistant: ["gpt", "assistant"], system: ["system"], tool: ["tool"] + roles: dict[str, list[str]] | None + # Whether to drop the system turn from the dataset. Only works with chat_template. + # This does not drop the default system message from chat_template if it exists. If + # you wish to, we recommend using a custom jinja template with the default system + # message removed or adding a system turn with empty content. + drop_system_message: bool | None + # Trust remote code for untrusted source + trust_remote_code: bool | None = False + # The specific revision of the dataset to use when loading from the Hugging Face Hub. + # This can be a commit hash, tag, or branch name. If not specified, the latest version + # will be used. This parameter is ignored for local datasets. + revision: str | None + + # For DPODataset: + path: str | None + split: str | None + type: UserDefinedDPOType | str | None + # For UserDefinedDPOType: + field_system: str | None + field_prompt: str | None + field_chosen: str | None + field_rejected: str | None + prompt_format: str | None + chosen_format: str | None + rejected_format: str | None + data_files: list[str] | None + revision: str | None + field_messages: str | None + + # For KTODataset: + path: str | None + split: str | None + type: UserDefinedKTOType | str | None + # For UserDefinedKTOType: + field_system: str | None + field_prompt: str | None + field_completion: str | None + field_label: bool | None + prompt_format: str | None + completion_format: str | None + data_files: list[str] | None + trust_remote_code: bool | None = False + revision: str | None + + # For StepwiseSupervisedDataset: + path: str | None + split: str | None + data_files: list[str] | None + revision: str | None + step_separator: str | None + max_completion_length: int | None + train_on_last_step_only: bool | None + +# A list of one or more datasets to eval the model with. You can use either +# test_datasets, or val_set_size, but not both. +test_datasets: Annotated[list[SFTDataset | DPODataset | KTODataset | StepwiseSupervisedDataset], MinLen(1)] | None + # For SFTDataset: + # HuggingFace dataset repo | s3:// | gs:// | path to local file or directory + path: str | None + # name of dataset split to load from + split: str | None + # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection] + type: str | UserDefinedPrompterType | None + # For UserDefinedPrompterType: + # Custom user instruction prompt + system_prompt: str | None + # Use {system} as key to be replaced + system_format: str | None + field_system: str | None + field_instruction: str | None + field_input: str | None + field_output: str | None + + # Customizable to be single line or multi-line. Use {instruction}/{input} as key to + # be replaced. 'format' can include {input} + format: str | None + # 'no_input_format' cannot include {input} + no_input_format: str | None + input_transform: str | None + # split dataset into N pieces (use with shards_idx) + shards: int | None + # the index of sharded dataset to use + shards_idx: int | None + # process dataset in N sequential chunks for memory efficiency (exclusive with + # `shards`) + preprocess_shards: int | None + conversation: str | None + + # The name of the chat template to use for training, following values are supported: + # tokenizer_default: Uses the chat template that is available in the + # tokenizer_config.json. If the chat template is not available in the tokenizer, it + # will raise an error. This is the default. + # alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates + # are available in the axolotl codebase at src/axolotl/utils/chat_templates.py. + # tokenizer_default_fallback_*: where * is the name of the chat template to fallback + # to if the tokenizer does not have a chat template else default to tokenizer. E.g. + # tokenizer_default_fallback_chatml. jinja: Uses a custom jinja template for the chat + # template. The custom jinja template should be provided in the chat_template_jinja + # field. + chat_template: ChatTemplate | str | None + # Custom jinja chat template or path to jinja file. Used only if `chat_template: + # jinja` or empty. + chat_template_jinja: str | None + # path to source data files + data_files: str | list[str] | None + input_format: str | None + # name of dataset configuration to load + name: str | None + # defines the datatype when path is a file + ds_type: str | None + # For `completion` datasets only, uses the provided field instead of `text` column + field: str | None + field_human: str | None + field_model: str | None + # Key containing the messages (default: "messages") + field_messages: str | None + # Key containing the tools (default: "tools"). Must be a list[dict] and follow [JSON + # schema](https://json-schema.org/learn/getting-started-step-by-step). + field_tools: str | None + # Key containing the reasoning trace (default: "reasoning_content"). + field_thinking: str | None + # The key the chat template expects that indicates the reasoning trace. + template_thinking_key: str | None + + message_field_role: str | None + + message_field_content: str | None + # Mapping of properties from the input dataset to the chat template. (default: + # message_property_mappings={'role':'role', 'content':'content'}) If a property exists + # in the template but not in this mapping, the system will attempt to load it directly + # from the message using the property name as the key. Example: In the mapping below, + # 'from' is loaded from input dataset and used as 'role', while 'value' is loaded and + # used as 'content' in the chat template. + message_property_mappings: dict[str, str] | None + # The key in the message turn that indicates via boolean whether tokens of a turn + # should be considered for training. Useful to selectively train on certain turns + # besides the `roles_to_train`. + message_field_training: str | None + # The key in the message turn that contains the training details. Useful to + # selectively train on certain tokens in a turn. The value of the key is a List[Dict] + # containing `begin_offset` (start character index in content), `end_offset` (end + # character index in content), and `train` (boolean whether to train). + message_field_training_detail: str | None + # (for Qwen3 template only) Whether to split the assistant content based on a + # reasoning trace inside delimited tags + split_thinking: bool | None + logprobs_field: str | None + temperature: float | None + # Roles to train on. The tokens from these roles will be considered for the loss. + roles_to_train: list[str] | None + # Which EOS tokens to train on in the conversation. Possible values are: all: train on + # all EOS tokens, turn (default): train on the EOS token at the end of each trainable + # turn, last: train on the last EOS token in the conversation + train_on_eos: Literal['all', 'turn', 'last'] | None + # Roles mapping in the messages. The format is {target_role: [source_roles]}. All + # source roles will be mapped to the target role. The default is: user: ["human", + # "user"], assistant: ["gpt", "assistant"], system: ["system"], tool: ["tool"] + roles: dict[str, list[str]] | None + # Whether to drop the system turn from the dataset. Only works with chat_template. + # This does not drop the default system message from chat_template if it exists. If + # you wish to, we recommend using a custom jinja template with the default system + # message removed or adding a system turn with empty content. + drop_system_message: bool | None + # Trust remote code for untrusted source + trust_remote_code: bool | None = False + # The specific revision of the dataset to use when loading from the Hugging Face Hub. + # This can be a commit hash, tag, or branch name. If not specified, the latest version + # will be used. This parameter is ignored for local datasets. + revision: str | None + + # For DPODataset: + path: str | None + split: str | None + type: UserDefinedDPOType | str | None + # For UserDefinedDPOType: + field_system: str | None + field_prompt: str | None + field_chosen: str | None + field_rejected: str | None + prompt_format: str | None + chosen_format: str | None + rejected_format: str | None + data_files: list[str] | None + revision: str | None + field_messages: str | None + + # For KTODataset: + path: str | None + split: str | None + type: UserDefinedKTOType | str | None + # For UserDefinedKTOType: + field_system: str | None + field_prompt: str | None + field_completion: str | None + field_label: bool | None + prompt_format: str | None + completion_format: str | None + data_files: list[str] | None + trust_remote_code: bool | None = False + revision: str | None + + # For StepwiseSupervisedDataset: + path: str | None + split: str | None + data_files: list[str] | None + revision: str | None + step_separator: str | None + max_completion_length: int | None + train_on_last_step_only: bool | None + +# If false, the datasets will not be shuffled and will keep their original order in +# `datasets`. The same applies to the `test_datasets` option and the +# `pretraining_dataset` option. Default is true. +shuffle_merged_datasets: bool | None = True +# If true, each dataset in `datasets` will be shuffled before merging. This allows +# curriculum learning strategies to be applied at the dataset level. Default is false. +shuffle_before_merging_datasets: bool | None = False +# Axolotl attempts to save the dataset as an arrow after packing the data together so +# subsequent training attempts load faster, relative path +dataset_prepared_path: str | None +# Num shards for whole dataset +dataset_shard_num: int | None +# Index of shard to use for whole dataset +dataset_shard_idx: int | None +skip_prepare_dataset: bool | None = False +# Number of shards to save the prepared dataset +num_dataset_shards_to_save: int | None + +# Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize +pretraining_dataset: Annotated[list[PretrainingDataset | SFTDataset], MinLen(1)] | None + # For PretrainingDataset: + name: str | None + path: str | None + split: str | None = train + text_column: str | None = text + type: str | None = pretrain + trust_remote_code: bool | None = False + data_files: str | None + skip: int | None + + # For SFTDataset: + # HuggingFace dataset repo | s3:// | gs:// | path to local file or directory + path: str | None + # name of dataset split to load from + split: str | None + # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection] + type: str | UserDefinedPrompterType | None + # For UserDefinedPrompterType: + # Custom user instruction prompt + system_prompt: str | None + # Use {system} as key to be replaced + system_format: str | None + field_system: str | None + field_instruction: str | None + field_input: str | None + field_output: str | None + + # Customizable to be single line or multi-line. Use {instruction}/{input} as key to + # be replaced. 'format' can include {input} + format: str | None + # 'no_input_format' cannot include {input} + no_input_format: str | None + input_transform: str | None + # split dataset into N pieces (use with shards_idx) + shards: int | None + # the index of sharded dataset to use + shards_idx: int | None + # process dataset in N sequential chunks for memory efficiency (exclusive with + # `shards`) + preprocess_shards: int | None + conversation: str | None + + # The name of the chat template to use for training, following values are supported: + # tokenizer_default: Uses the chat template that is available in the + # tokenizer_config.json. If the chat template is not available in the tokenizer, it + # will raise an error. This is the default. + # alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates + # are available in the axolotl codebase at src/axolotl/utils/chat_templates.py. + # tokenizer_default_fallback_*: where * is the name of the chat template to fallback + # to if the tokenizer does not have a chat template else default to tokenizer. E.g. + # tokenizer_default_fallback_chatml. jinja: Uses a custom jinja template for the chat + # template. The custom jinja template should be provided in the chat_template_jinja + # field. + chat_template: ChatTemplate | str | None + # Custom jinja chat template or path to jinja file. Used only if `chat_template: + # jinja` or empty. + chat_template_jinja: str | None + # path to source data files + data_files: str | list[str] | None + input_format: str | None + # name of dataset configuration to load + name: str | None + # defines the datatype when path is a file + ds_type: str | None + # For `completion` datasets only, uses the provided field instead of `text` column + field: str | None + field_human: str | None + field_model: str | None + # Key containing the messages (default: "messages") + field_messages: str | None + # Key containing the tools (default: "tools"). Must be a list[dict] and follow [JSON + # schema](https://json-schema.org/learn/getting-started-step-by-step). + field_tools: str | None + # Key containing the reasoning trace (default: "reasoning_content"). + field_thinking: str | None + # The key the chat template expects that indicates the reasoning trace. + template_thinking_key: str | None + + message_field_role: str | None + + message_field_content: str | None + # Mapping of properties from the input dataset to the chat template. (default: + # message_property_mappings={'role':'role', 'content':'content'}) If a property exists + # in the template but not in this mapping, the system will attempt to load it directly + # from the message using the property name as the key. Example: In the mapping below, + # 'from' is loaded from input dataset and used as 'role', while 'value' is loaded and + # used as 'content' in the chat template. + message_property_mappings: dict[str, str] | None + # The key in the message turn that indicates via boolean whether tokens of a turn + # should be considered for training. Useful to selectively train on certain turns + # besides the `roles_to_train`. + message_field_training: str | None + # The key in the message turn that contains the training details. Useful to + # selectively train on certain tokens in a turn. The value of the key is a List[Dict] + # containing `begin_offset` (start character index in content), `end_offset` (end + # character index in content), and `train` (boolean whether to train). + message_field_training_detail: str | None + # (for Qwen3 template only) Whether to split the assistant content based on a + # reasoning trace inside delimited tags + split_thinking: bool | None + logprobs_field: str | None + temperature: float | None + # Roles to train on. The tokens from these roles will be considered for the loss. + roles_to_train: list[str] | None + # Which EOS tokens to train on in the conversation. Possible values are: all: train on + # all EOS tokens, turn (default): train on the EOS token at the end of each trainable + # turn, last: train on the last EOS token in the conversation + train_on_eos: Literal['all', 'turn', 'last'] | None + # Roles mapping in the messages. The format is {target_role: [source_roles]}. All + # source roles will be mapped to the target role. The default is: user: ["human", + # "user"], assistant: ["gpt", "assistant"], system: ["system"], tool: ["tool"] + roles: dict[str, list[str]] | None + # Whether to drop the system turn from the dataset. Only works with chat_template. + # This does not drop the default system message from chat_template if it exists. If + # you wish to, we recommend using a custom jinja template with the default system + # message removed or adding a system turn with empty content. + drop_system_message: bool | None + # Trust remote code for untrusted source + trust_remote_code: bool | None = False + # The specific revision of the dataset to use when loading from the Hugging Face Hub. + # This can be a commit hash, tag, or branch name. If not specified, the latest version + # will be used. This parameter is ignored for local datasets. + revision: str | None + +# The maximum number of processes to use while preprocessing your input dataset. This +# defaults to `os.cpu_count()` if not set. For Runpod VMs, it will default to number of +# vCPUs via RUNPOD_CPU_COUNT. +dataset_processes: int | None +# Deduplicates datasets and test_datasets with identical entries +dataset_exact_deduplication: bool | None +# Keep dataset in memory while preprocessing. Only needed if cached dataset is taking +# too much storage +dataset_keep_in_memory: bool | None +dataloader_pin_memory: bool | None +dataloader_num_workers: int | None +dataloader_prefetch_factor: int | None +dataloader_drop_last: bool | None + +accelerator_config: dict[str, Any] | None + +remove_unused_columns: bool | None + +# Push prepared dataset to hub - repo_org/repo_name +push_dataset_to_hub: str | None +# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private +# datasets. Required to be true when used in combination with `push_dataset_to_hub` +hf_use_auth_token: bool | None + +device: Any | None +# Passed through to transformers when loading the model when launched without +# accelerate. Use `sequential` when training w/ model parallelism to limit memory +device_map: Any | None +world_size: int | None +# Don't mess with this, it's here for accelerate and torchrun +local_rank: int | None +ddp: bool | None + +# Seed for reproducibility +seed: int | None +# Advanced DDP Arguments - timeout +ddp_timeout: int | None +# Advanced DDP Arguments - bucket cap in MB +ddp_bucket_cap_mb: int | None +# Advanced DDP Arguments - broadcast buffers +ddp_broadcast_buffers: bool | None +ddp_find_unused_parameters: bool | None + +# Approximate number of predictions sent to wandb depending on batch size. Enabled above +# 0. Default is 0 +eval_table_size: int | None +# Total number of tokens generated for predictions sent to wandb. Default is 128 +eval_max_new_tokens: int | None +# Whether to run causal language model evaluation for metrics in +# `eval_causal_lm_metrics` +do_causal_lm_eval: bool | None +# HF evaluate metrics used during evaluation. Default is ['sacrebleu', 'comet', 'ter', +# 'chrf', 'perplexity'] +eval_causal_lm_metrics: list[str] | None +do_bench_eval: bool | None +bench_dataset: str | None +bench_split: str | None +metric_for_best_model: str | None +greater_is_better: bool | None + +# High loss value, indicating the learning has broken down (a good estimate is ~2 times +# the loss at the start of training) +loss_watchdog_threshold: float | None +# Number of high-loss steps in a row before the trainer aborts (default: 3) +loss_watchdog_patience: int | None + +# Run garbage collection every `gc_steps` steps. -1 will run on epoch end and before +# evaluations. Default is 0 (disabled). +gc_steps: int | None + +# Use CUDA bf16. bool or 'full' for `bf16_full_eval`, or 'auto' for automatic detection. +# require >=ampere +bf16: Literal['auto'] | bool | None = auto +# Use CUDA fp16 +fp16: bool | None +# Enable FP8 mixed precision training using TorchAO. Best used in combination with +# torch.compile. +fp8: bool | None +# Enable FSDP float8 all-gather optimization for FP8 training. Can improve training +# speed by 10-15% when FSDP is enabled. +fp8_enable_fsdp_float8_all_gather: bool | None +# No AMP (automatic mixed precision) - require >=ampere +bfloat16: bool | None +# No AMP (automatic mixed precision) +float16: bool | None +# Use CUDA tf32 - require >=ampere +tf32: bool | None +float32: bool | None + +# Whether to use gradient checkpointing. Available options are: true, false, 'offload', +# 'offload_disk'. +# https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing +gradient_checkpointing: Literal['offload', 'offload_disk'] | bool | None = False +# Additional kwargs to pass to the trainer for gradient checkpointing +gradient_checkpointing_kwargs: dict[str, Any] | None +# Whether to offload activations. Available options are: true, false, 'legacy', 'disk'. +activation_offloading: Literal['legacy', 'disk'] | bool | None = False + +unfrozen_parameters: list[str] | None + +# The maximum length of an input to train with, this should typically be less than 2048 +# as most models have a token/context limit of 2048 +sequence_len: int = 512 +# The maximum length of an input for evaluation. If not specified, defaults to +# sequence_len +eval_sequence_len: int | None +min_sample_len: int | None +# maximum prompt length for RL training +max_prompt_len: int = 512 +# Use efficient multi-packing with block diagonal attention and per sequence +# position_ids. Recommend set to 'true' +sample_packing: bool | None +# The number of samples packed at a time. Increasing the following values helps with +# packing, but usually only slightly (<%1.) +sample_packing_group_size: int | None = 100000 +# The number of samples which can be packed into one sequence. Increase if using a large +# sequence_len with many short samples. +sample_packing_bin_size: int | None = 200 +# Whether to pack samples sequentially +sample_packing_sequentially: bool | None +# The multiprocessing start method to use for packing. Should be 'fork', 'spawn' or +# 'forkserver' +sample_packing_mp_start_method: str | None +# Set to 'false' if getting errors during eval with sample_packing on +eval_sample_packing: bool | None +# Pad inputs so each step uses constant sized buffers. This will reduce memory +# fragmentation and may prevent OOMs, by re-using memory more efficiently. Defaults to +# True if `sample_packing` enabled +pad_to_sequence_len: bool | None +# Whether to use sequential sampling for curriculum learning +curriculum_sampling: bool | None +multipack_real_batches: bool | None +# whether to concatenate samples during pretraining +pretraining_sample_concatenation: bool | None -# Whether to use xformers attention patch https://github.com/facebookresearch/xformers -xformers_attention: bool | None -# Whether to use scaled-dot-product attention https://pytorch.org/docs/stable/generated/ -# torch.nn.functional.scaled_dot_product_attention.html -sdp_attention: bool | None -# Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf -s2_attention: bool | None -flex_attention: bool | None -flex_attn_compile_kwargs: dict[str, Any] | None -# Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention -flash_attention: bool | None -# Whether to use flash-attention cross entropy implementation - advanced use only -flash_attn_cross_entropy: bool | None -# Whether to use flash-attention rms norm implementation - advanced use only -flash_attn_rms_norm: bool | None -# Whether to fuse part of the MLP into a single operation -flash_attn_fuse_mlp: bool | None -# Whether to use bettertransformers -flash_optimum: bool | None - -eager_attention: bool | None - -# Specify a custom attention implementation, used mostly for kernels. -attn_implementation: str | None - -unsloth_cross_entropy_loss: bool | None -unsloth_lora_mlp: bool | None -unsloth_lora_qkv: bool | None -unsloth_lora_o: bool | None -unsloth_rms_norm: bool | None -unsloth_rope: bool | None +# Use batch flattening for speedups when not using sample_packing +batch_flattening: Literal['auto'] | bool | None + +use_pose: bool | None +pose_split_on_token_ids: list[int] | None +pose_max_context_len: int | None +pose_num_chunks: int | None + +pretrain_multipack_buffer_size: int | None = 10000 +# whether to prevent cross attention for packed sequences during pretraining +pretrain_multipack_attn: bool | None = True + +# Whether to use xformers attention patch https://github.com/facebookresearch/xformers +xformers_attention: bool | None +# Whether to use scaled-dot-product attention https://pytorch.org/docs/stable/generated/ +# torch.nn.functional.scaled_dot_product_attention.html +sdp_attention: bool | None +# Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf +s2_attention: bool | None +flex_attention: bool | None +flex_attn_compile_kwargs: dict[str, Any] | None +# Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention +flash_attention: bool | None +# Whether to use flash-attention cross entropy implementation - advanced use only +flash_attn_cross_entropy: bool | None +# Whether to use flash-attention rms norm implementation - advanced use only +flash_attn_rms_norm: bool | None +# Whether to fuse part of the MLP into a single operation +flash_attn_fuse_mlp: bool | None +# Whether to use bettertransformers +flash_optimum: bool | None -# Apply custom LoRA autograd functions and activation function Triton kernels for speed -# and memory savings. See: https://docs.axolotl.ai/docs/lora_optims.html -lora_mlp_kernel: bool | None -# Apply custom LoRA autograd functions and activation function Triton kernels for speed -# and memory savings. See: https://docs.axolotl.ai/docs/lora_optims.html -lora_qkv_kernel: bool | None -# Apply custom LoRA autograd functions and activation function Triton kernels for speed -# and memory savings. See: https://docs.axolotl.ai/docs/lora_optims.html -lora_o_kernel: bool | None - -# Whether to use chunked cross entropy loss for memory efficiency -chunked_cross_entropy: bool | None -# Number of chunks to use for chunked cross entropy loss -chunked_cross_entropy_num_chunks: int | None - -# Whether to use ALST tiled mlp for memory efficient long context -tiled_mlp: bool | None - -# Number of shards to use for ALST tiled mlp. If unset, it will be set based on -# seqlen/hidden_size -tiled_mlp_num_shards: int | None +eager_attention: bool | None + +# Specify a custom attention implementation, used mostly for kernels. +attn_implementation: str | None + +unsloth_cross_entropy_loss: bool | None +unsloth_lora_mlp: bool | None +unsloth_lora_qkv: bool | None +unsloth_lora_o: bool | None +unsloth_rms_norm: bool | None +unsloth_rope: bool | None + +# Apply custom LoRA autograd functions and activation function Triton kernels for speed +# and memory savings. See: https://docs.axolotl.ai/docs/lora_optims.html +lora_mlp_kernel: bool | None +# Apply custom LoRA autograd functions and activation function Triton kernels for speed +# and memory savings. See: https://docs.axolotl.ai/docs/lora_optims.html +lora_qkv_kernel: bool | None +# Apply custom LoRA autograd functions and activation function Triton kernels for speed +# and memory savings. See: https://docs.axolotl.ai/docs/lora_optims.html +lora_o_kernel: bool | None -# Whether to use original mlp for ALST tiled mlp. Otherwise uses a generic MLP based on -# llama. -tiled_mlp_use_original_mlp: bool | None = True - -llama4_linearized_experts: bool | None - -# Deepspeed config path. e.g., deepspeed_configs/zero3.json -deepspeed: str | dict[str, Any] | None -# Whether to use deepcompile for faster training with deepspeed -deepcompile: bool | None -# FSDP configuration -fsdp: list[str] | None - -# FSDP configuration options -fsdp_config: dict[str, Any] | None -# FSDP version -fsdp_version: int | None -fsdp_final_state_dict_type: Literal['FULL_STATE_DICT', 'LOCAL_STATE_DICT', 'SHARDED_STATE_DICT'] | None - -# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for -# no eval. -val_set_size: float | None = 0.0 - -# Number of devices to shard across. If not set, will use all available devices. -dp_shard_size: int | None -# Number of devices to replicate across. -dp_replicate_size: int | None -# Deprecated: use `context_parallel_size` instead -sequence_parallel_degree: int | None -# Set to a divisor of the number of GPUs available to split sequences into chunks of -# equal size. Use in long context training to prevent OOM when sequences cannot fit into -# a single GPU's VRAM. E.g., if 4 GPUs are available, set this value to 2 to split each -# sequence into two equal-sized subsequences, or set to 4 to split into four equal-sized -# subsequences. See https://docs.axolotl.ai/docs/sequence_parallelism.html for more -# details. -context_parallel_size: int | None -# Optional; strides across the key dimension. Larger values use more memory but should -# make training faster. Must evenly divide the number of KV heads in your model. -heads_k_stride: int | None -# One of 'varlen_llama3', 'batch_ring', 'batch_zigzag', 'batch_stripe'. Defaults to -# 'varlen_llama3' in the sample packing case, and 'batch_ring' in the non-sample packing -# case. -ring_attn_func: RingAttnFunc | None -# Number of tensor parallel processes in TP group. Only supported with DeepSpeed AutoTP. -tensor_parallel_size: int | None - -# Add or change special tokens. If you add tokens here, you don't need to add them to -# the `tokens` list. -special_tokens: SpecialTokensConfig | None - # For SpecialTokensConfig: - bos_token: str | None - eos_token: str | None - pad_token: str | None - unk_token: str | None - additional_special_tokens: list[str] | None - -# Add extra tokens to the tokenizer -tokens: list[str] | None -# Mapping token_id to new_token_string to override reserved added_tokens in the -# tokenizer. Only works for tokens that are not part of the base vocab (aka are -# added_tokens). Can be checked if they exist in tokenizer.json added_tokens. -added_tokens_overrides: dict[int, str] | None - -# Whether to use torch.compile and which backend to use. setting to `auto` will enable -# torch compile when torch>=2.6.0 -torch_compile: Literal['auto'] | bool | None -# Backend to use for torch.compile -torch_compile_backend: str | None -torch_compile_mode: Literal['default', 'reduce-overhead', 'max-autotune'] | None - -# Maximum number of iterations to train for. It precedes num_epochs which means that if -# both are set, num_epochs will not be guaranteed. e.g., when 1 epoch is 1000 steps => -# `num_epochs: 2` and `max_steps: 100` will train for 100 steps -max_steps: int | None -# Number of warmup steps. Cannot use with warmup_ratio -warmup_steps: int | None -# Warmup ratio. Cannot use with warmup_steps -warmup_ratio: float | None -# Leave empty to eval at each epoch, integer for every N steps. float for fraction of -# total steps -eval_steps: int | float | None -# Number of times per epoch to run evals, mutually exclusive with eval_steps -evals_per_epoch: int | None -# Set to `no` to skip evaluation, `epoch` at end of each epoch, leave empty to infer -# from `eval_steps` -eval_strategy: str | None - -# Leave empty to save at each epoch, integer for every N steps. float for fraction of -# total steps -save_steps: int | float | None -# Number of times per epoch to save a checkpoint, mutually exclusive with save_steps -saves_per_epoch: int | None -# Set to `no` to skip checkpoint saves, `epoch` at end of each epoch, `best` when better -# result is achieved, leave empty to infer from `save_steps` -save_strategy: str | None -# Checkpoints saved at a time -save_total_limit: int | None -# Whether to checkpoint a model after the first step of training. Defaults to False. -save_first_step: bool | None - -# Logging frequency -logging_steps: int | None -# Stop training after this many evaluation losses have increased in a row. https://huggi -# ngface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppin -# gCallback -early_stopping_patience: int | None -load_best_model_at_end: bool | None = False -# Save only the model weights, skipping the optimizer. Using this means you can't resume -# from checkpoints. -save_only_model: bool | None = False -# Use tensorboard for logging -use_tensorboard: bool | None -# Enable the pytorch profiler to capture the first N steps of training to the -# output_dir. see https://pytorch.org/blog/understanding-gpu-memory-1/ for more -# information. Snapshots can be visualized @ https://pytorch.org/memory_viz -profiler_steps: int | None -# Which step to start the profiler at. Useful for only capturing a few steps mid-run. -profiler_steps_start: int | None = 0 -# bool of whether to include tokens trainer per second in the training metrics. This -# iterates over the entire dataset once, so it takes some time. -include_tokens_per_second: bool | None - -# NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to -# add noise to embeddings. Currently only supported on Llama and Mistral -neftune_noise_alpha: float | None - -# Parameter controlling the relative ratio loss weight in the ORPO loss. Passed to -# `beta` in `ORPOConfig` due to trl mapping. -orpo_alpha: float | None -# Weighting of NLL term in loss from RPO paper -rpo_alpha: float | None -# Target reward margin for the SimPO loss -simpo_gamma: float | None -# Weight of the BC regularizer -cpo_alpha: float | None - -# Factor for desirable loss term in KTO loss -kto_desirable_weight: float | None -# Factor for undesirable loss term in KTO loss -kto_undesirable_weight: float | None -# The beta parameter for the RL training -rl_beta: float | None - -# Defines the max memory usage per gpu on the system. Passed through to transformers -# when loading the model. -max_memory: dict[int | Literal['cpu', 'disk'], int | str] | None -# Limit the memory for all available GPUs to this amount (if an integer, expressed in -# gigabytes); default: unset -gpu_memory_limit: int | str | None -# Whether to use low_cpu_mem_usage -low_cpu_mem_usage: bool | None - -# The name of the chat template to use for training, following values are supported: -# tokenizer_default: Uses the chat template that is available in the -# tokenizer_config.json. If the chat template is not available in the tokenizer, it will -# raise an error. This is the default value. -# alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates -# are available in the axolotl codebase at src/axolotl/utils/chat_templates.py. -# tokenizer_default_fallback_*: where * is the name of the chat template to fallback to. -# E.g. tokenizer_default_fallback_chatml. This is useful when the chat template is not -# available in the tokenizer. jinja: Uses a custom jinja template for the chat template. -# The custom jinja template should be provided in the chat_template_jinja field. The -# selected chat template will be saved to the tokenizer_config.json for easier -# inferencing -chat_template: ChatTemplate | Annotated[str, StringConstraints(pattern='^tokenizer_default_fallback_')] | None -# Custom jinja template or path to jinja file for chat template. This will be only used -# if chat_template is set to `jinja` or `null` (in which case chat_template is -# automatically set to `jinja`). Default is null. -chat_template_jinja: str | None -# Additional kwargs to pass to the chat template. This is useful for customizing the -# chat template. For example, you can pass `thinking=False` to add a generation prompt -# to the chat template. -chat_template_kwargs: dict[str, Any] | None -# Custom EOT (End-of-Turn) tokens to mask/unmask during training. These tokens mark the -# boundaries between conversation turns. For example: ['/INST', '</s>', -# '[/SYSTEM_PROMPT]']. If not specified, defaults to just the model's eos_token. This is -# useful for templates that use multiple delimiter tokens. -eot_tokens: list[str] | None -# Changes the default system message. Currently only supports chatml. -default_system_message: str | None - -fix_untrained_tokens: int | list[int] | None - -is_preprocess: bool | None -preprocess_iterable: bool | None - -# Total number of tokens - internal use -total_num_tokens: int | None -total_supervised_tokens: int | None -# You can set these packing optimizations AFTER starting a training at least once. The -# trainer will provide recommended values for these values. -sample_packing_eff_est: float | None -axolotl_config_path: str | None - -# Internal use only - Used to identify which the model is based on -is_falcon_derived_model: bool | None -# Internal use only - Used to identify which the model is based on -is_llama_derived_model: bool | None -# Internal use only - Used to identify which the model is based on. Please note that if -# you set this to true, `padding_side` will be set to 'left' by default -is_mistral_derived_model: bool | None -# Internal use only - Used to identify which the model is based on -is_qwen_derived_model: bool | None - -# Add plugins to extend the pipeline. See `src/axolotl/integrations` for the available -# plugins or doc below for more details. -# https://docs.axolotl.ai/docs/custom_integrations.html -plugins: list[str] | None - -# This is the huggingface model that contains *.pt, *.safetensors, or *.bin files. This -# can also be a relative path to a model on disk -base_model: str (required) -# If the base_model repo on hf hub doesn't include configuration .json files, You can -# set that here, or leave this empty to default to base_model -base_model_config: str | None -cls_model_config: str | None -# Optional tokenizer configuration path in case you want to use a different tokenizer -# than the one defined in the base model -tokenizer_config: str | None -# use_fast option for tokenizer loading from_pretrained, default to True -tokenizer_use_fast: bool | None -# Whether to use the legacy tokenizer setting, defaults to True -tokenizer_legacy: bool | None -# Whether to use mistral-common tokenizer. If set to True, it will use the mistral- -# common tokenizer. -tokenizer_use_mistral_common: bool | None -# Corresponding tokenizer for the model AutoTokenizer is a good choice -tokenizer_type: str | None -# transformers processor class -processor_type: str | None -# Trust remote code for untrusted source -trust_remote_code: bool | None - -# Don't move the model to the device before sharding. This is an experimental feature -# that may be included in the future as the default. -experimental_skip_move_to_device: bool | None - -# Use custom kernels, e.g. MegaBlocks. -use_kernels: bool | None - -# Model loading quantization config -model_quantization_config: Literal['Mxfp4Config'] | None -# kwargs for model quantization config -model_quantization_config_kwargs: dict[str, Any] | None +# Whether to use chunked cross entropy loss for memory efficiency +chunked_cross_entropy: bool | None +# Number of chunks to use for chunked cross entropy loss +chunked_cross_entropy_num_chunks: int | None + +# Whether to use ALST tiled mlp for memory efficient long context +tiled_mlp: bool | None + +# Number of shards to use for ALST tiled mlp. If unset, it will be set based on +# seqlen/hidden_size +tiled_mlp_num_shards: int | None + +# Whether to use original mlp for ALST tiled mlp. Otherwise uses a generic MLP based on +# llama. +tiled_mlp_use_original_mlp: bool | None = True + +llama4_linearized_experts: bool | None + +# Deepspeed config path. e.g., deepspeed_configs/zero3.json +deepspeed: str | dict[str, Any] | None +# Whether to use deepcompile for faster training with deepspeed +deepcompile: bool | None +# FSDP configuration +fsdp: list[str] | None + +# FSDP configuration options +fsdp_config: dict[str, Any] | None +# FSDP version +fsdp_version: int | None +fsdp_final_state_dict_type: Literal['FULL_STATE_DICT', 'LOCAL_STATE_DICT', 'SHARDED_STATE_DICT'] | None + +# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for +# no eval. +val_set_size: float | None = 0.0 + +# Number of devices to shard across. If not set, will use all available devices. +dp_shard_size: int | None +# Number of devices to replicate across. +dp_replicate_size: int | None +# Deprecated: use `context_parallel_size` instead +sequence_parallel_degree: int | None +# Set to a divisor of the number of GPUs available to split sequences into chunks of +# equal size. Use in long context training to prevent OOM when sequences cannot fit into +# a single GPU's VRAM. E.g., if 4 GPUs are available, set this value to 2 to split each +# sequence into two equal-sized subsequences, or set to 4 to split into four equal-sized +# subsequences. See https://docs.axolotl.ai/docs/sequence_parallelism.html for more +# details. +context_parallel_size: int | None +# Optional; strides across the key dimension. Larger values use more memory but should +# make training faster. Must evenly divide the number of KV heads in your model. +heads_k_stride: int | None +# One of 'varlen_llama3', 'batch_ring', 'batch_zigzag', 'batch_stripe'. Defaults to +# 'varlen_llama3' in the sample packing case, and 'batch_ring' in the non-sample packing +# case. +ring_attn_func: RingAttnFunc | None +# Number of tensor parallel processes in TP group. Only supported with DeepSpeed AutoTP. +tensor_parallel_size: int | None + +# Add or change special tokens. If you add tokens here, you don't need to add them to +# the `tokens` list. +special_tokens: SpecialTokensConfig | None + # For SpecialTokensConfig: + bos_token: str | None + eos_token: str | None + pad_token: str | None + unk_token: str | None + additional_special_tokens: list[str] | None + +# Add extra tokens to the tokenizer +tokens: list[str] | None +# Mapping token_id to new_token_string to override reserved added_tokens in the +# tokenizer. Only works for tokens that are not part of the base vocab (aka are +# added_tokens). Can be checked if they exist in tokenizer.json added_tokens. +added_tokens_overrides: dict[int, str] | None + +# Whether to use torch.compile and which backend to use. setting to `auto` will enable +# torch compile when torch>=2.6.0 +torch_compile: Literal['auto'] | bool | None +# Backend to use for torch.compile +torch_compile_backend: str | None +torch_compile_mode: Literal['default', 'reduce-overhead', 'max-autotune'] | None + +# Maximum number of iterations to train for. It precedes num_epochs which means that if +# both are set, num_epochs will not be guaranteed. e.g., when 1 epoch is 1000 steps => +# `num_epochs: 2` and `max_steps: 100` will train for 100 steps +max_steps: int | None +# Number of warmup steps. Cannot use with warmup_ratio +warmup_steps: int | None +# Warmup ratio. Cannot use with warmup_steps +warmup_ratio: float | None +# Leave empty to eval at each epoch, integer for every N steps. float for fraction of +# total steps +eval_steps: int | float | None +# Number of times per epoch to run evals, mutually exclusive with eval_steps +evals_per_epoch: int | None +# Set to `no` to skip evaluation, `epoch` at end of each epoch, leave empty to infer +# from `eval_steps` +eval_strategy: str | None + +# Leave empty to save at each epoch, integer for every N steps. float for fraction of +# total steps +save_steps: int | float | None +# Number of times per epoch to save a checkpoint, mutually exclusive with save_steps +saves_per_epoch: int | None +# Set to `no` to skip checkpoint saves, `epoch` at end of each epoch, `best` when better +# result is achieved, leave empty to infer from `save_steps` +save_strategy: str | None +# Checkpoints saved at a time +save_total_limit: int | None +# Whether to checkpoint a model after the first step of training. Defaults to False. +save_first_step: bool | None + +# Logging frequency +logging_steps: int | None +# Stop training after this many evaluation losses have increased in a row. https://huggi +# ngface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppin +# gCallback +early_stopping_patience: int | None +load_best_model_at_end: bool | None = False +# Save only the model weights, skipping the optimizer. Using this means you can't resume +# from checkpoints. +save_only_model: bool | None = False +# Use tensorboard for logging +use_tensorboard: bool | None +# Enable the pytorch profiler to capture the first N steps of training to the +# output_dir. see https://pytorch.org/blog/understanding-gpu-memory-1/ for more +# information. Snapshots can be visualized @ https://pytorch.org/memory_viz +profiler_steps: int | None +# Which step to start the profiler at. Useful for only capturing a few steps mid-run. +profiler_steps_start: int | None = 0 +# bool of whether to include tokens trainer per second in the training metrics. This +# iterates over the entire dataset once, so it takes some time. +include_tokens_per_second: bool | None + +# NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to +# add noise to embeddings. Currently only supported on Llama and Mistral +neftune_noise_alpha: float | None + +# Parameter controlling the relative ratio loss weight in the ORPO loss. Passed to +# `beta` in `ORPOConfig` due to trl mapping. +orpo_alpha: float | None +# Weighting of NLL term in loss from RPO paper +rpo_alpha: float | None +# Target reward margin for the SimPO loss +simpo_gamma: float | None +# Weight of the BC regularizer +cpo_alpha: float | None + +# Factor for desirable loss term in KTO loss +kto_desirable_weight: float | None +# Factor for undesirable loss term in KTO loss +kto_undesirable_weight: float | None +# The beta parameter for the RL training +rl_beta: float | None + +# Defines the max memory usage per gpu on the system. Passed through to transformers +# when loading the model. +max_memory: dict[int | Literal['cpu', 'disk'], int | str] | None +# Limit the memory for all available GPUs to this amount (if an integer, expressed in +# gigabytes); default: unset +gpu_memory_limit: int | str | None +# Whether to use low_cpu_mem_usage +low_cpu_mem_usage: bool | None + +# The name of the chat template to use for training, following values are supported: +# tokenizer_default: Uses the chat template that is available in the +# tokenizer_config.json. If the chat template is not available in the tokenizer, it will +# raise an error. This is the default value. +# alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates +# are available in the axolotl codebase at src/axolotl/utils/chat_templates.py. +# tokenizer_default_fallback_*: where * is the name of the chat template to fallback to. +# E.g. tokenizer_default_fallback_chatml. This is useful when the chat template is not +# available in the tokenizer. jinja: Uses a custom jinja template for the chat template. +# The custom jinja template should be provided in the chat_template_jinja field. The +# selected chat template will be saved to the tokenizer_config.json for easier +# inferencing +chat_template: ChatTemplate | Annotated[str, StringConstraints(pattern='^tokenizer_default_fallback_')] | None +# Custom jinja template or path to jinja file for chat template. This will be only used +# if chat_template is set to `jinja` or `null` (in which case chat_template is +# automatically set to `jinja`). Default is null. +chat_template_jinja: str | None +# Additional kwargs to pass to the chat template. This is useful for customizing the +# chat template. For example, you can pass `thinking=False` to add a generation prompt +# to the chat template. +chat_template_kwargs: dict[str, Any] | None +# Custom EOT (End-of-Turn) tokens to mask/unmask during training. These tokens mark the +# boundaries between conversation turns. For example: ['/INST', '</s>', +# '[/SYSTEM_PROMPT]']. If not specified, defaults to just the model's eos_token. This is +# useful for templates that use multiple delimiter tokens. +eot_tokens: list[str] | None +# Changes the default system message. Currently only supports chatml. +default_system_message: str | None + +fix_untrained_tokens: int | list[int] | None + +is_preprocess: bool | None +preprocess_iterable: bool | None + +# Total number of tokens - internal use +total_num_tokens: int | None +total_supervised_tokens: int | None +# You can set these packing optimizations AFTER starting a training at least once. The +# trainer will provide recommended values for these values. +sample_packing_eff_est: float | None +axolotl_config_path: str | None + +# Internal use only - Used to identify which the model is based on +is_falcon_derived_model: bool | None +# Internal use only - Used to identify which the model is based on +is_llama_derived_model: bool | None +# Internal use only - Used to identify which the model is based on. Please note that if +# you set this to true, `padding_side` will be set to 'left' by default +is_mistral_derived_model: bool | None +# Internal use only - Used to identify which the model is based on +is_qwen_derived_model: bool | None + +# Add plugins to extend the pipeline. See `src/axolotl/integrations` for the available +# plugins or doc below for more details. +# https://docs.axolotl.ai/docs/custom_integrations.html +plugins: list[str] | None + +# This is the huggingface model that contains *.pt, *.safetensors, or *.bin files. This +# can also be a relative path to a model on disk +base_model: str (required) +# If the base_model repo on hf hub doesn't include configuration .json files, You can +# set that here, or leave this empty to default to base_model +base_model_config: str | None +cls_model_config: str | None +# Optional tokenizer configuration path in case you want to use a different tokenizer +# than the one defined in the base model +tokenizer_config: str | None +# use_fast option for tokenizer loading from_pretrained, default to True +tokenizer_use_fast: bool | None +# Whether to use the legacy tokenizer setting, defaults to True +tokenizer_legacy: bool | None +# Whether to use mistral-common tokenizer. If set to True, it will use the mistral- +# common tokenizer. +tokenizer_use_mistral_common: bool | None +# Corresponding tokenizer for the model AutoTokenizer is a good choice +tokenizer_type: str | None +# transformers processor class +processor_type: str | None +# Trust remote code for untrusted source +trust_remote_code: bool | None -# Where to save the full-finetuned model to -output_dir: str = ./model-out -# push checkpoints to hub -hub_model_id: str | None -# how to push checkpoints to hub -hub_strategy: str | None -# Save model as safetensors (require safetensors package). Default True -save_safetensors: bool | None = True - -# This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer -load_in_8bit: bool | None = False -# Use bitsandbytes 4 bit -load_in_4bit: bool | None = False - -# If you want to use 'lora' or 'qlora' or leave blank to train all parameters in -# original model -adapter: str | None -# If you already have a lora model trained that you want to load, put that here. This -# means after training, if you want to test the model, you should set this to the value -# of `output_dir`. Note that if you merge an adapter to the base model, a new -# subdirectory `merged` will be created under the `output_dir`. -lora_model_dir: str | None -lora_r: int | None -lora_alpha: int | None -lora_fan_in_fan_out: bool | None -lora_target_modules: str | list[str] | None -lora_target_parameters: str | list[str] | None -# If true, will target all linear modules -lora_target_linear: bool | None -# If you added new tokens to the tokenizer, you may need to save some LoRA modules -# because they need to know the new tokens. For LLaMA and Mistral, you need to save -# `embed_tokens` and `lm_head`. It may vary for other models. `embed_tokens` converts -# tokens to embeddings, and `lm_head` converts embeddings to token probabilities. -lora_modules_to_save: list[str] | None -lora_dropout: float | None = 0.0 -# The layer indices to transform, otherwise, apply to all layers -peft_layers_to_transform: list[int] | None -peft_layers_pattern: list[str] | None - -peft: PeftConfig | None - # For PeftConfig: - # Configuration options for loftq initialization for LoRA - loftq_config: LoftQConfig | None - # For LoftQConfig: - # typically 4 bits - loftq_bits: int = 4 - -# Whether to use DoRA. -peft_use_dora: bool | None -# Whether to use RSLoRA. -peft_use_rslora: bool | None -# List of layer indices to replicate. -peft_layer_replication: list[tuple[int, int]] | None -# How to initialize LoRA weights. Default to True which is MS original implementation. -peft_init_lora_weights: bool | str | None - -# load qlora model in sharded format for FSDP using answer.ai technique. -qlora_sharded_model_loading: bool | None = False -# Do the LoRA/PEFT loading on CPU -- this is required if the base model is so large it -# takes up most or all of the available GPU VRAM, e.g. during a model and LoRA merge -lora_on_cpu: bool | None -# Whether you are training a 4-bit GPTQ quantized model -gptq: bool | None -# optional overrides to the bnb 4bit quantization configuration -bnb_config_kwargs: dict[str, Any] | None - -# loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4. -loraplus_lr_ratio: float | None -# loraplus learning rate for lora embedding layers. Default value is 1e-6. -loraplus_lr_embedding: float | None = 1e-06 - -merge_lora: bool | None - -# Whether to use ReLoRA. Use with jagged_restart_*steps options. -relora: bool | None -# threshold for optimizer magnitude when pruning -relora_prune_ratio: float | None -# True to perform lora weight merges on cpu during restarts, for modest gpu memory -# savings -relora_cpu_offload: bool | None - -# how often to reset for jagged restarts -jagged_restart_steps: int | None -# how many warmup steps to take after reset for jagged restarts -jagged_restart_warmup_steps: int | None -# how many anneal steps to take before reset for jagged restarts -jagged_restart_anneal_steps: int | None - -# If greater than 1, backpropagation will be skipped and the gradients will be -# accumulated for the given number of steps. -gradient_accumulation_steps: int | None = 1 -# The number of samples to include in each batch. This is the number of samples sent to -# each GPU. Batch size per gpu = micro_batch_size * gradient_accumulation_steps -micro_batch_size: int | None = 1 -# Total batch size, we do not recommended setting this manually -batch_size: int | None -# per gpu micro batch size for evals, defaults to value of micro_batch_size -eval_batch_size: int | None - -# whether to find batch size that fits in memory. Passed to underlying transformers -# Trainer -auto_find_batch_size: bool | None - -# Whether to mask out or include the human's prompt from the training labels -train_on_inputs: bool | None = False -# Group similarly sized data to minimize padding. May be slower to start, as it must -# download and sort the entire dataset. Note that training loss may have an oscillating -# pattern with this enabled. -group_by_length: bool | None - -learning_rate: str | float (required) -embedding_lr: float | None -embedding_lr_scale: float | None -# Specify weight decay -weight_decay: float | None = 0.0 -# Specify optimizer -optimizer: OptimizerNames | CustomSupportedOptimizers | None = OptimizerNames.ADAMW_TORCH_FUSED -# Dictionary of arguments to pass to the optimizer -optim_args: str | dict[str, Any] | None -# The target modules to optimize, i.e. the module names that you would like to train, -# right now this is used only for GaLore algorithm -optim_target_modules: list[str] | Literal['all_linear'] | None -# Path to torch distx for optim 'adamw_anyprecision' -torchdistx_path: str | None -lr_scheduler: SchedulerType | Literal['one_cycle'] | Literal['rex'] | None = SchedulerType.COSINE -# Specify a scheduler and kwargs to use with the optimizer -lr_scheduler_kwargs: dict[str, Any] | None -lr_quadratic_warmup: bool | None -# decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of -# peak lr -cosine_min_lr_ratio: float | None -# freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means -# start cosine_min_lr at 80% of training step -cosine_constant_lr_ratio: float | None -# Learning rate div factor -lr_div_factor: float | None - -lr_groups: list[LrGroup] | None - # For LrGroup: - name: str (required) - modules: list[str] (required) - lr: float (required) - -# adamw hyperparams -adam_epsilon: float | None -# only used for CAME Optimizer -adam_epsilon2: float | None -# adamw hyperparams -adam_beta1: float | None -# adamw hyperparams -adam_beta2: float | None -# only used for CAME Optimizer -adam_beta3: float | None - -# Dion Optimizer learning rate -dion_lr: float | None -# Dion Optimizer momentum -dion_momentum: float | None -# Dion Optimizer: r/d fraction for low-rank approximation. Used to compute the low-rank -# dimension. -dion_rank_fraction: float | None = 1.0 -# Dion Optimizer: Round up the low-rank dimension to a multiple of this number. This may -# be useful to ensure even sharding. -dion_rank_multiple_of: int | None = 1 - -# Gradient clipping max norm -max_grad_norm: float | None -num_epochs: float = 1.0 - -use_wandb: bool | None -# Set the name of your wandb run -wandb_name: str | None -# Set the ID of your wandb run -wandb_run_id: str | None -# "offline" to save run metadata locally and not sync to the server, "disabled" to turn -# off wandb -wandb_mode: str | None -# Your wandb project name -wandb_project: str | None -# A wandb Team name if using a Team -wandb_entity: str | None -wandb_watch: str | None -# "checkpoint" to log model to wandb Artifacts every `save_steps` or "end" to log only -# at the end of training -wandb_log_model: str | None - -use_mlflow: bool | None -# URI to mlflow -mlflow_tracking_uri: str | None -# Your experiment name -mlflow_experiment_name: str | None -# Your run name -mlflow_run_name: str | None -# set to true to copy each saved checkpoint on each save to mlflow artifact registry -hf_mlflow_log_artifacts: bool | None - -# Enable or disable Comet integration. -use_comet: bool | None -# API key for Comet. Recommended to set via `comet login`. -comet_api_key: str | None -# Workspace name in Comet. Defaults to the user's default workspace. -comet_workspace: str | None -# Project name in Comet. Defaults to Uncategorized. -comet_project_name: str | None -# Identifier for the experiment. Used to append data to an existing experiment or -# control the key of new experiments. Default to a random key. -comet_experiment_key: str | None -# Create a new experiment ("create") or log to an existing one ("get"). Default -# ("get_or_create") auto-selects based on configuration. -comet_mode: str | None -# Set to True to log data to Comet server, or False for offline storage. Default is -# True. -comet_online: bool | None -# Dictionary for additional configuration settings, see the doc for more details. -comet_experiment_config: dict[str, Any] | None - -# the number of activate layers in LISA -lisa_n_layers: int | None -# how often to switch layers in LISA -lisa_step_interval: int | None -# path under the model to access the layers -lisa_layers_attribute: str | None = model.layers - -gradio_title: str | None -gradio_share: bool | None -gradio_server_name: str | None -gradio_server_port: int | None -gradio_max_new_tokens: int | None -gradio_temperature: float | None - -use_ray: bool = False -ray_run_name: str | None -ray_num_workers: int = 1 -resources_per_worker: dict +# Don't move the model to the device before sharding. This is an experimental feature +# that may be included in the future as the default. +experimental_skip_move_to_device: bool | None + +# Use custom kernels, e.g. MegaBlocks. +use_kernels: bool | None + +# Model loading quantization config +model_quantization_config: Literal['Mxfp4Config'] | None +# kwargs for model quantization config +model_quantization_config_kwargs: dict[str, Any] | None + +# Where to save the full-finetuned model to +output_dir: str = ./model-out +# push checkpoints to hub +hub_model_id: str | None +# how to push checkpoints to hub +hub_strategy: str | None +# Save model as safetensors (require safetensors package). Default True +save_safetensors: bool | None = True + +# This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer +load_in_8bit: bool | None = False +# Use bitsandbytes 4 bit +load_in_4bit: bool | None = False + +# If you want to use 'lora' or 'qlora' or leave blank to train all parameters in +# original model +adapter: str | None +# If you already have a lora model trained that you want to load, put that here. This +# means after training, if you want to test the model, you should set this to the value +# of `output_dir`. Note that if you merge an adapter to the base model, a new +# subdirectory `merged` will be created under the `output_dir`. +lora_model_dir: str | None +lora_r: int | None +lora_alpha: int | None +lora_fan_in_fan_out: bool | None +lora_target_modules: str | list[str] | None +lora_target_parameters: str | list[str] | None +# If true, will target all linear modules +lora_target_linear: bool | None +# If you added new tokens to the tokenizer, you may need to save some LoRA modules +# because they need to know the new tokens. For LLaMA and Mistral, you need to save +# `embed_tokens` and `lm_head`. It may vary for other models. `embed_tokens` converts +# tokens to embeddings, and `lm_head` converts embeddings to token probabilities. +lora_modules_to_save: list[str] | None +lora_dropout: float | None = 0.0 +# The layer indices to transform, otherwise, apply to all layers +peft_layers_to_transform: list[int] | None +peft_layers_pattern: list[str] | None + +peft: PeftConfig | None + # For PeftConfig: + # Configuration options for loftq initialization for LoRA + loftq_config: LoftQConfig | None + # For LoftQConfig: + # typically 4 bits + loftq_bits: int = 4 + +# Whether to use DoRA. +peft_use_dora: bool | None +# Whether to use RSLoRA. +peft_use_rslora: bool | None +# List of layer indices to replicate. +peft_layer_replication: list[tuple[int, int]] | None +# How to initialize LoRA weights. Default to True which is MS original implementation. +peft_init_lora_weights: bool | str | None + +# load qlora model in sharded format for FSDP using answer.ai technique. +qlora_sharded_model_loading: bool | None = False +# Do the LoRA/PEFT loading on CPU -- this is required if the base model is so large it +# takes up most or all of the available GPU VRAM, e.g. during a model and LoRA merge +lora_on_cpu: bool | None +# Whether you are training a 4-bit GPTQ quantized model +gptq: bool | None +# optional overrides to the bnb 4bit quantization configuration +bnb_config_kwargs: dict[str, Any] | None + +# loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4. +loraplus_lr_ratio: float | None +# loraplus learning rate for lora embedding layers. Default value is 1e-6. +loraplus_lr_embedding: float | None = 1e-06 + +merge_lora: bool | None + +# Whether to use ReLoRA. Use with jagged_restart_*steps options. +relora: bool | None +# threshold for optimizer magnitude when pruning +relora_prune_ratio: float | None +# True to perform lora weight merges on cpu during restarts, for modest gpu memory +# savings +relora_cpu_offload: bool | None + +# how often to reset for jagged restarts +jagged_restart_steps: int | None +# how many warmup steps to take after reset for jagged restarts +jagged_restart_warmup_steps: int | None +# how many anneal steps to take before reset for jagged restarts +jagged_restart_anneal_steps: int | None + +# If greater than 1, backpropagation will be skipped and the gradients will be +# accumulated for the given number of steps. +gradient_accumulation_steps: int | None = 1 +# The number of samples to include in each batch. This is the number of samples sent to +# each GPU. Batch size per gpu = micro_batch_size * gradient_accumulation_steps +micro_batch_size: int | None = 1 +# Total batch size, we do not recommended setting this manually +batch_size: int | None +# per gpu micro batch size for evals, defaults to value of micro_batch_size +eval_batch_size: int | None + +# whether to find batch size that fits in memory. Passed to underlying transformers +# Trainer +auto_find_batch_size: bool | None + +# Whether to mask out or include the human's prompt from the training labels +train_on_inputs: bool | None = False +# Group similarly sized data to minimize padding. May be slower to start, as it must +# download and sort the entire dataset. Note that training loss may have an oscillating +# pattern with this enabled. +group_by_length: bool | None + +learning_rate: str | float (required) +embedding_lr: float | None +embedding_lr_scale: float | None +# Specify weight decay +weight_decay: float | None = 0.0 +# Specify optimizer +optimizer: OptimizerNames | CustomSupportedOptimizers | None = OptimizerNames.ADAMW_TORCH_FUSED +# Dictionary of arguments to pass to the optimizer +optim_args: str | dict[str, Any] | None +# The target modules to optimize, i.e. the module names that you would like to train, +# right now this is used only for GaLore algorithm +optim_target_modules: list[str] | Literal['all_linear'] | None +# Path to torch distx for optim 'adamw_anyprecision' +torchdistx_path: str | None +lr_scheduler: SchedulerType | Literal['one_cycle'] | Literal['rex'] | None = SchedulerType.COSINE +# Specify a scheduler and kwargs to use with the optimizer +lr_scheduler_kwargs: dict[str, Any] | None +lr_quadratic_warmup: bool | None +# decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of +# peak lr +cosine_min_lr_ratio: float | None +# freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means +# start cosine_min_lr at 80% of training step +cosine_constant_lr_ratio: float | None +# Learning rate div factor +lr_div_factor: float | None + +lr_groups: list[LrGroup] | None + # For LrGroup: + name: str (required) + modules: list[str] (required) + lr: float (required) + +# adamw hyperparams +adam_epsilon: float | None +# only used for CAME Optimizer +adam_epsilon2: float | None +# adamw hyperparams +adam_beta1: float | None +# adamw hyperparams +adam_beta2: float | None +# only used for CAME Optimizer +adam_beta3: float | None + +# Dion Optimizer learning rate +dion_lr: float | None +# Dion Optimizer momentum +dion_momentum: float | None +# Dion Optimizer: r/d fraction for low-rank approximation. Used to compute the low-rank +# dimension. +dion_rank_fraction: float | None = 1.0 +# Dion Optimizer: Round up the low-rank dimension to a multiple of this number. This may +# be useful to ensure even sharding. +dion_rank_multiple_of: int | None = 1 + +# Gradient clipping max norm +max_grad_norm: float | None +num_epochs: float = 1.0 + +use_wandb: bool | None +# Set the name of your wandb run +wandb_name: str | None +# Set the ID of your wandb run +wandb_run_id: str | None +# "offline" to save run metadata locally and not sync to the server, "disabled" to turn +# off wandb +wandb_mode: str | None +# Your wandb project name +wandb_project: str | None +# A wandb Team name if using a Team +wandb_entity: str | None +wandb_watch: str | None +# "checkpoint" to log model to wandb Artifacts every `save_steps` or "end" to log only +# at the end of training +wandb_log_model: str | None + +use_mlflow: bool | None +# URI to mlflow +mlflow_tracking_uri: str | None +# Your experiment name +mlflow_experiment_name: str | None +# Your run name +mlflow_run_name: str | None +# set to true to copy each saved checkpoint on each save to mlflow artifact registry +hf_mlflow_log_artifacts: bool | None + +# Enable or disable Comet integration. +use_comet: bool | None +# API key for Comet. Recommended to set via `comet login`. +comet_api_key: str | None +# Workspace name in Comet. Defaults to the user's default workspace. +comet_workspace: str | None +# Project name in Comet. Defaults to Uncategorized. +comet_project_name: str | None +# Identifier for the experiment. Used to append data to an existing experiment or +# control the key of new experiments. Default to a random key. +comet_experiment_key: str | None +# Create a new experiment ("create") or log to an existing one ("get"). Default +# ("get_or_create") auto-selects based on configuration. +comet_mode: str | None +# Set to True to log data to Comet server, or False for offline storage. Default is +# True. +comet_online: bool | None +# Dictionary for additional configuration settings, see the doc for more details. +comet_experiment_config: dict[str, Any] | None + +# the number of activate layers in LISA +lisa_n_layers: int | None +# how often to switch layers in LISA +lisa_step_interval: int | None +# path under the model to access the layers +lisa_layers_attribute: str | None = model.layers -# The size of the image to resize to. It can be an integer (resized into padded-square -# image) or a tuple (width, height).If not provided, we will attempt to load from -# preprocessor.size, otherwise, images won't be resized. -image_size: int | tuple[int, int] | None -# The resampling algorithm to use for image resizing. Default is bilinear. Please refer -# to PIL.Image.Resampling for more details. -image_resize_algorithm: Literal['bilinear', 'bicubic', 'lanczos'] | Resampling | None - -# optional overrides to the base model configuration -overrides_of_model_config: dict[str, Any] | None -# optional overrides the base model loading from_pretrained -overrides_of_model_kwargs: dict[str, Any] | None -# If you want to specify the type of model to load, AutoModelForCausalLM is a good -# choice too -type_of_model: str | None -# You can specify to choose a specific model revision from huggingface hub -revision_of_model: str | None - -max_packed_sequence_len: int | None -rope_scaling: Any | None -noisy_embedding_alpha: float | None -dpo_beta: float | None -evaluation_strategy: str | None +gradio_title: str | None +gradio_share: bool | None +gradio_server_name: str | None +gradio_server_port: int | None +gradio_max_new_tokens: int | None +gradio_temperature: float | None + +use_ray: bool = False +ray_run_name: str | None +ray_num_workers: int = 1 +resources_per_worker: dict + +# The size of the image to resize to. It can be an integer (resized into padded-square +# image) or a tuple (width, height).If not provided, we will attempt to load from +# preprocessor.size, otherwise, images won't be resized. +image_size: int | tuple[int, int] | None +# The resampling algorithm to use for image resizing. Default is bilinear. Please refer +# to PIL.Image.Resampling for more details. +image_resize_algorithm: Literal['bilinear', 'bicubic', 'lanczos'] | Resampling | None + +# optional overrides to the base model configuration +overrides_of_model_config: dict[str, Any] | None +# optional overrides the base model loading from_pretrained +overrides_of_model_kwargs: dict[str, Any] | None +# If you want to specify the type of model to load, AutoModelForCausalLM is a good +# choice too +type_of_model: str | None +# You can specify to choose a specific model revision from huggingface hub +revision_of_model: str | None + +max_packed_sequence_len: int | None +rope_scaling: Any | None +noisy_embedding_alpha: float | None +dpo_beta: float | None +evaluation_strategy: str | None diff --git a/docs/dataset-formats/conversation.html b/docs/dataset-formats/conversation.html index 5b8b2ee48..afff7df0f 100644 --- a/docs/dataset-formats/conversation.html +++ b/docs/dataset-formats/conversation.html @@ -775,9 +775,10 @@ Note

Tools need to follow JSON schema.

+

Example config for Llama4:

chat_template: llama4
 datasets:
-  - path: ...
+  - path: Nanobit/text-tools-2k-test
     type: chat_template
     # field_tools: tools # default is `tools`
diff --git a/search.json b/search.json index e06a05930..064bb0fda 100644 --- a/search.json +++ b/search.json @@ -970,14 +970,14 @@ "href": "docs/api/prompt_strategies.chat_template.html", "title": "prompt_strategies.chat_template", "section": "", - "text": "prompt_strategies.chat_template\nHF Chat Templates prompt strategy\n\n\n\n\n\nName\nDescription\n\n\n\n\nChatTemplatePrompter\nPrompter for HF chat templates\n\n\nChatTemplateStrategy\nTokenizing strategy for instruction-based prompts.\n\n\nMistralPrompter\nMistral prompter for chat template.\n\n\nMistralStrategy\nMistral strategy for chat template.\n\n\nStrategyLoader\nLoad chat template strategy based on configuration.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplatePrompter(\n tokenizer,\n chat_template,\n processor=None,\n max_length=2048,\n message_property_mappings=None,\n message_field_training=None,\n message_field_training_detail=None,\n field_messages='messages',\n field_system='system',\n field_tools='tools',\n roles=None,\n chat_template_kwargs=None,\n drop_system_message=False,\n)\nPrompter for HF chat templates\n\n\n\n\n\nName\nDescription\n\n\n\n\nbuild_prompt\nBuild a prompt from a conversation.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplatePrompter.build_prompt(\n conversation,\n add_generation_prompt=False,\n images=None,\n tools=None,\n)\nBuild a prompt from a conversation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconversation\nlist[dict]\nA list of messages.\nrequired\n\n\nadd_generation_prompt\n\nWhether to add a generation prompt.\nFalse\n\n\nimages\n\nA list of images. (optional)\nNone\n\n\ntools\n\nA list of tools. (optional)\nNone\n\n\n\n\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy(\n prompter,\n tokenizer,\n train_on_inputs,\n sequence_len,\n roles_to_train=None,\n train_on_eos=None,\n train_on_eot=None,\n eot_tokens=None,\n split_thinking=False,\n)\nTokenizing strategy for instruction-based prompts.\n\n\n\n\n\nName\nDescription\n\n\n\n\nfind_first_eot_token\nFind the first EOT token in the input_ids starting from start_idx.\n\n\nfind_turn\nLocate the starting and ending indices of the specified turn in a conversation.\n\n\ntokenize_prompt\nPublic method that can handle either a single prompt or a batch of prompts.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.find_first_eot_token(\n input_ids,\n start_idx,\n)\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.find_turn(\n turns,\n turn_idx,\n tools=None,\n)\nLocate the starting and ending indices of the specified turn in a conversation.\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.tokenize_prompt(prompt)\nPublic method that can handle either a single prompt or a batch of prompts.\n\n\n\n\n\nprompt_strategies.chat_template.MistralPrompter(*args, **kwargs)\nMistral prompter for chat template.\n\n\n\nprompt_strategies.chat_template.MistralStrategy(\n prompter,\n tokenizer,\n train_on_inputs,\n sequence_len,\n roles_to_train=None,\n train_on_eos=None,\n train_on_eot=None,\n eot_tokens=None,\n split_thinking=False,\n)\nMistral strategy for chat template.\n\n\n\n\n\nName\nDescription\n\n\n\n\nfind_first_eot_token\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\n\n\nprompt_strategies.chat_template.MistralStrategy.find_first_eot_token(\n input_ids,\n start_idx,\n)\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\n\n\nprompt_strategies.chat_template.StrategyLoader()\nLoad chat template strategy based on configuration." + "text": "prompt_strategies.chat_template\nHF Chat Templates prompt strategy\n\n\n\n\n\nName\nDescription\n\n\n\n\nChatTemplatePrompter\nPrompter for HF chat templates\n\n\nChatTemplateStrategy\nTokenizing strategy for instruction-based prompts.\n\n\nMistralPrompter\nMistral prompter for chat template.\n\n\nMistralStrategy\nMistral strategy for chat template.\n\n\nStrategyLoader\nLoad chat template strategy based on configuration.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplatePrompter(\n tokenizer,\n chat_template,\n processor=None,\n max_length=2048,\n message_property_mappings=None,\n message_field_training=None,\n message_field_training_detail=None,\n field_messages='messages',\n field_system='system',\n field_tools='tools',\n field_thinking='reasoning_content',\n roles=None,\n template_thinking_key='reasoning_content',\n chat_template_kwargs=None,\n drop_system_message=False,\n)\nPrompter for HF chat templates\n\n\n\n\n\nName\nDescription\n\n\n\n\nbuild_prompt\nBuild a prompt from a conversation.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplatePrompter.build_prompt(\n conversation,\n add_generation_prompt=False,\n images=None,\n tools=None,\n)\nBuild a prompt from a conversation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconversation\nlist[dict]\nA list of messages.\nrequired\n\n\nadd_generation_prompt\n\nWhether to add a generation prompt.\nFalse\n\n\nimages\n\nA list of images. (optional)\nNone\n\n\ntools\n\nA list of tools. (optional)\nNone\n\n\n\n\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy(\n prompter,\n tokenizer,\n train_on_inputs,\n sequence_len,\n roles_to_train=None,\n train_on_eos=None,\n train_on_eot=None,\n eot_tokens=None,\n split_thinking=False,\n)\nTokenizing strategy for instruction-based prompts.\n\n\n\n\n\nName\nDescription\n\n\n\n\nfind_first_eot_token\nFind the first EOT token in the input_ids starting from start_idx.\n\n\nfind_turn\nLocate the starting and ending indices of the specified turn in a conversation.\n\n\ntokenize_prompt\nPublic method that can handle either a single prompt or a batch of prompts.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.find_first_eot_token(\n input_ids,\n start_idx,\n)\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.find_turn(\n turns,\n turn_idx,\n tools=None,\n)\nLocate the starting and ending indices of the specified turn in a conversation.\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.tokenize_prompt(prompt)\nPublic method that can handle either a single prompt or a batch of prompts.\n\n\n\n\n\nprompt_strategies.chat_template.MistralPrompter(*args, **kwargs)\nMistral prompter for chat template.\n\n\n\nprompt_strategies.chat_template.MistralStrategy(\n prompter,\n tokenizer,\n train_on_inputs,\n sequence_len,\n roles_to_train=None,\n train_on_eos=None,\n train_on_eot=None,\n eot_tokens=None,\n split_thinking=False,\n)\nMistral strategy for chat template.\n\n\n\n\n\nName\nDescription\n\n\n\n\nfind_first_eot_token\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\n\n\nprompt_strategies.chat_template.MistralStrategy.find_first_eot_token(\n input_ids,\n start_idx,\n)\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\n\n\nprompt_strategies.chat_template.StrategyLoader()\nLoad chat template strategy based on configuration." }, { "objectID": "docs/api/prompt_strategies.chat_template.html#classes", "href": "docs/api/prompt_strategies.chat_template.html#classes", "title": "prompt_strategies.chat_template", "section": "", - "text": "Name\nDescription\n\n\n\n\nChatTemplatePrompter\nPrompter for HF chat templates\n\n\nChatTemplateStrategy\nTokenizing strategy for instruction-based prompts.\n\n\nMistralPrompter\nMistral prompter for chat template.\n\n\nMistralStrategy\nMistral strategy for chat template.\n\n\nStrategyLoader\nLoad chat template strategy based on configuration.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplatePrompter(\n tokenizer,\n chat_template,\n processor=None,\n max_length=2048,\n message_property_mappings=None,\n message_field_training=None,\n message_field_training_detail=None,\n field_messages='messages',\n field_system='system',\n field_tools='tools',\n roles=None,\n chat_template_kwargs=None,\n drop_system_message=False,\n)\nPrompter for HF chat templates\n\n\n\n\n\nName\nDescription\n\n\n\n\nbuild_prompt\nBuild a prompt from a conversation.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplatePrompter.build_prompt(\n conversation,\n add_generation_prompt=False,\n images=None,\n tools=None,\n)\nBuild a prompt from a conversation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconversation\nlist[dict]\nA list of messages.\nrequired\n\n\nadd_generation_prompt\n\nWhether to add a generation prompt.\nFalse\n\n\nimages\n\nA list of images. (optional)\nNone\n\n\ntools\n\nA list of tools. (optional)\nNone\n\n\n\n\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy(\n prompter,\n tokenizer,\n train_on_inputs,\n sequence_len,\n roles_to_train=None,\n train_on_eos=None,\n train_on_eot=None,\n eot_tokens=None,\n split_thinking=False,\n)\nTokenizing strategy for instruction-based prompts.\n\n\n\n\n\nName\nDescription\n\n\n\n\nfind_first_eot_token\nFind the first EOT token in the input_ids starting from start_idx.\n\n\nfind_turn\nLocate the starting and ending indices of the specified turn in a conversation.\n\n\ntokenize_prompt\nPublic method that can handle either a single prompt or a batch of prompts.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.find_first_eot_token(\n input_ids,\n start_idx,\n)\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.find_turn(\n turns,\n turn_idx,\n tools=None,\n)\nLocate the starting and ending indices of the specified turn in a conversation.\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.tokenize_prompt(prompt)\nPublic method that can handle either a single prompt or a batch of prompts.\n\n\n\n\n\nprompt_strategies.chat_template.MistralPrompter(*args, **kwargs)\nMistral prompter for chat template.\n\n\n\nprompt_strategies.chat_template.MistralStrategy(\n prompter,\n tokenizer,\n train_on_inputs,\n sequence_len,\n roles_to_train=None,\n train_on_eos=None,\n train_on_eot=None,\n eot_tokens=None,\n split_thinking=False,\n)\nMistral strategy for chat template.\n\n\n\n\n\nName\nDescription\n\n\n\n\nfind_first_eot_token\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\n\n\nprompt_strategies.chat_template.MistralStrategy.find_first_eot_token(\n input_ids,\n start_idx,\n)\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\n\n\nprompt_strategies.chat_template.StrategyLoader()\nLoad chat template strategy based on configuration." + "text": "Name\nDescription\n\n\n\n\nChatTemplatePrompter\nPrompter for HF chat templates\n\n\nChatTemplateStrategy\nTokenizing strategy for instruction-based prompts.\n\n\nMistralPrompter\nMistral prompter for chat template.\n\n\nMistralStrategy\nMistral strategy for chat template.\n\n\nStrategyLoader\nLoad chat template strategy based on configuration.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplatePrompter(\n tokenizer,\n chat_template,\n processor=None,\n max_length=2048,\n message_property_mappings=None,\n message_field_training=None,\n message_field_training_detail=None,\n field_messages='messages',\n field_system='system',\n field_tools='tools',\n field_thinking='reasoning_content',\n roles=None,\n template_thinking_key='reasoning_content',\n chat_template_kwargs=None,\n drop_system_message=False,\n)\nPrompter for HF chat templates\n\n\n\n\n\nName\nDescription\n\n\n\n\nbuild_prompt\nBuild a prompt from a conversation.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplatePrompter.build_prompt(\n conversation,\n add_generation_prompt=False,\n images=None,\n tools=None,\n)\nBuild a prompt from a conversation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nconversation\nlist[dict]\nA list of messages.\nrequired\n\n\nadd_generation_prompt\n\nWhether to add a generation prompt.\nFalse\n\n\nimages\n\nA list of images. (optional)\nNone\n\n\ntools\n\nA list of tools. (optional)\nNone\n\n\n\n\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy(\n prompter,\n tokenizer,\n train_on_inputs,\n sequence_len,\n roles_to_train=None,\n train_on_eos=None,\n train_on_eot=None,\n eot_tokens=None,\n split_thinking=False,\n)\nTokenizing strategy for instruction-based prompts.\n\n\n\n\n\nName\nDescription\n\n\n\n\nfind_first_eot_token\nFind the first EOT token in the input_ids starting from start_idx.\n\n\nfind_turn\nLocate the starting and ending indices of the specified turn in a conversation.\n\n\ntokenize_prompt\nPublic method that can handle either a single prompt or a batch of prompts.\n\n\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.find_first_eot_token(\n input_ids,\n start_idx,\n)\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.find_turn(\n turns,\n turn_idx,\n tools=None,\n)\nLocate the starting and ending indices of the specified turn in a conversation.\n\n\n\nprompt_strategies.chat_template.ChatTemplateStrategy.tokenize_prompt(prompt)\nPublic method that can handle either a single prompt or a batch of prompts.\n\n\n\n\n\nprompt_strategies.chat_template.MistralPrompter(*args, **kwargs)\nMistral prompter for chat template.\n\n\n\nprompt_strategies.chat_template.MistralStrategy(\n prompter,\n tokenizer,\n train_on_inputs,\n sequence_len,\n roles_to_train=None,\n train_on_eos=None,\n train_on_eot=None,\n eot_tokens=None,\n split_thinking=False,\n)\nMistral strategy for chat template.\n\n\n\n\n\nName\nDescription\n\n\n\n\nfind_first_eot_token\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\n\n\nprompt_strategies.chat_template.MistralStrategy.find_first_eot_token(\n input_ids,\n start_idx,\n)\nFind the first EOT token in the input_ids starting from start_idx.\n\n\n\n\n\nprompt_strategies.chat_template.StrategyLoader()\nLoad chat template strategy based on configuration." }, { "objectID": "docs/api/core.training_args.html", @@ -1306,7 +1306,7 @@ "href": "docs/config-reference.html", "title": "Config Reference", "section": "", - "text": "# Allow overwrite yml config using from cli\nstrict: bool | None = False\n# Resume from a specific checkpoint dir\nresume_from_checkpoint: str | None\n# If resume_from_checkpoint isn't set and you simply want it to start where it left off.\n# Be careful with this being turned on between different models.\nauto_resume_from_checkpoints: bool | None\n# Resize the model embeddings when new tokens are added to multiples of 32. This is\n# reported to improve training speed on some models\nresize_token_embeddings_to_32x: bool | None\nmean_resizing_embeddings: bool | None = False\n\n# Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.\nshrink_embeddings: bool | None\n# Don't upcast the embeddings to float32 when using PEFT. Useful for low-VRAM GPUs\nembeddings_skip_upcast: bool | None\n\n# module to custom trainer class to use for training\ntrainer_cls: str | None\n\n# Use RL training: 'dpo', 'ipo', 'kto', 'simpo', 'orpo', 'grpo'\nrl: RLType | None\n\ntrl: TRLConfig | None\n # For TRLConfig:\n # Beta parameter for the RL training. Same as `rl_beta`. Use\n beta: float | None\n # Maximum length of the completion for RL training.\n max_completion_length: int | None\n\n # Whether to use VLLM for RL training.\n use_vllm: bool = False\n # VLLM mode to use, one of 'server' or 'colocate'\n vllm_mode: Literal['server', 'colocate'] | None\n # Host of the vLLM server to connect to.\n vllm_server_host: str | None = 0.0.0.0\n # Port of the vLLM server to connect to.\n vllm_server_port: int | None = 8000\n # Total timeout (in seconds) to wait for the vLLM server to respond.\n vllm_server_timeout: int | None\n # Regex for vLLM guided decoding.\n vllm_guided_decoding_regex: str | None\n\n # List of reward functions to load. Paths must be importable from current dir.\n reward_funcs: list[str] | None\n # List of reward weights for the reward functions.\n reward_weights: list[float] | None\n # Number of generations to sample.\n num_generations: int | None\n # Whether to log completions.\n log_completions: bool | None = False\n # Number of completions to print when log_completions is True.\n num_completions_to_print: int | None\n # Controls whether importance sampling ratios are computed at the `'token'` or\n # `'sequence'` level. For GSPO, use `sequence`, default is None which corresponds to\n # the original GRPO paper.\n importance_sampling_level: Literal['sequence', 'token'] | None\n\n # Whether to sync the reference model.\n sync_ref_model: bool | None = False\n # Mixup alpha for the reference model.\n ref_model_mixup_alpha: float | None = 0.9\n # Sync steps for the reference model.\n ref_model_sync_steps: int | None = 64\n # Whether to scale rewards by their standard deviation.\n scale_rewards: bool = True\n\n # Sampling temperature for the GRPO policy.\n temperature: float | None\n # Top-p sampling probability for the generation policy.\n top_p: float | None\n # Top-k sampling for the generation policy.\n top_k: int | None\n # Minimum probability for the generation policy.\n min_p: float | None\n # Penalty for tokens that appear in prompt and generated text.\n repetition_penalty: float | None\n # Number of iterations per batch (μ) for GRPO.\n num_iterations: int | None\n # Epsilon value for clipping in the GRPO algorithm.\n epsilon: float | None\n # Upper-bound epsilon value for clipping in the GRPO algorithm.\n epsilon_high: float | None\n # Whether to use Liger loss for GRPO.\n use_liger_loss: bool | None\n # Loss formulation to use. Supported values: grpo, bnpo, dr_grpo.\n loss_type: str | None\n # Whether to exclude truncated completions from loss calculation.\n mask_truncated_completions: bool = False\n\nvllm: VllmConfig | None\n # For VllmConfig:\n # Device to use for VLLM\n device: str | None = auto\n # Tensor parallel size for VLLM\n tensor_parallel_size: int | None\n # Data parallel size for VLLM\n data_parallel_size: int | None\n # GPU memory utilization for VLLM\n gpu_memory_utilization: float | None = 0.9\n # Data type for VLLM\n dtype: str | None = auto\n # Maximum length of the model context for VLLM\n max_model_len: int | None\n # Enable prefix caching for VLLM\n enable_prefix_caching: bool | None\n # Host for the vLLM server to start on\n host: str | None = 0.0.0.0\n # Port of the vLLM server to start on\n port: int | None = 8000\n\n # Enable reasoning for VLLM\n enable_reasoning: bool | None\n # Reasoning parser for VLLM\n reasoning_parser: str | None\n\nqat: QATConfig | None\n # For QATConfig:\n # Fake quantization layout to use for activation quantization. Valid options are\n # \"int4\" and \"int8\"\n activation_dtype: TorchIntDType | None\n # Fake quantization layout to use for weight quantization. Valid options are \"int4\"\n # and \"int8\"\n weight_dtype: TorchIntDType = TorchIntDType.int8\n # Quantize embedding\n quantize_embedding: bool | None = False\n # The number of elements in each group for per-group fake quantization\n group_size: int | None = 32\n # The number of steps to apply fake quantization after\n fake_quant_after_n_steps: int | None\n\nquantization: PTQConfig | None\n # For PTQConfig:\n # Fake quantization layout to use for weight quantization. Valid options are uintX for\n # X in [1, 2, 3, 4, 5, 6, 7], or int4, or int8\n weight_dtype: TorchIntDType = TorchIntDType.int8\n # Fake quantization layout to use for activation quantization. Valid options are\n # \"int4\" and \"int8\"\n activation_dtype: TorchIntDType | None\n # Whether to quantize the embedding layer.\n quantize_embedding: bool | None\n # The number of elements in each group for per-group fake quantization\n group_size: int | None = 32\n\n# Reward modelling: `True` or `False`\nreward_model: bool | None\n# Process reward modelling: `True` or `False`\nprocess_reward_model: bool | None\nnum_labels: int | None\n\n# Whether to perform weighting in DPO trainer\ndpo_use_weighting: bool | None\ndpo_use_logits_to_keep: bool | None\ndpo_label_smoothing: float | None\ndpo_norm_loss: bool | None\ndpo_padding_free: bool | None\ndpo_generate_during_eval: bool | None\n\n# A list of one or more datasets to finetune the model with\ndatasets: Annotated[list[SFTDataset | DPODataset | KTODataset | StepwiseSupervisedDataset], MinLen(1)] | None\n # For SFTDataset:\n # HuggingFace dataset repo | s3:// | gs:// | path to local file or directory\n path: str | None\n # name of dataset split to load from\n split: str | None\n # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]\n type: str | UserDefinedPrompterType | None\n # For UserDefinedPrompterType:\n # Custom user instruction prompt\n system_prompt: str | None\n # Use {system} as key to be replaced\n system_format: str | None\n field_system: str | None\n field_instruction: str | None\n field_input: str | None\n field_output: str | None\n\n # Customizable to be single line or multi-line. Use {instruction}/{input} as key to\n # be replaced. 'format' can include {input}\n format: str | None\n # 'no_input_format' cannot include {input}\n no_input_format: str | None\n input_transform: str | None\n # split dataset into N pieces (use with shards_idx)\n shards: int | None\n # the index of sharded dataset to use\n shards_idx: int | None\n # process dataset in N sequential chunks for memory efficiency (exclusive with\n # `shards`)\n preprocess_shards: int | None\n conversation: str | None\n\n # The name of the chat template to use for training, following values are supported:\n # tokenizer_default: Uses the chat template that is available in the\n # tokenizer_config.json. If the chat template is not available in the tokenizer, it\n # will raise an error. This is the default.\n # alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates\n # are available in the axolotl codebase at src/axolotl/utils/chat_templates.py.\n # tokenizer_default_fallback_*: where * is the name of the chat template to fallback\n # to if the tokenizer does not have a chat template else default to tokenizer. E.g.\n # tokenizer_default_fallback_chatml. jinja: Uses a custom jinja template for the chat\n # template. The custom jinja template should be provided in the chat_template_jinja\n # field.\n chat_template: ChatTemplate | str | None\n # Custom jinja chat template or path to jinja file. Used only if `chat_template:\n # jinja` or empty.\n chat_template_jinja: str | None\n # path to source data files\n data_files: str | list[str] | None\n input_format: str | None\n # name of dataset configuration to load\n name: str | None\n # defines the datatype when path is a file\n ds_type: str | None\n # For `completion` datasets only, uses the provided field instead of `text` column\n field: str | None\n field_human: str | None\n field_model: str | None\n # Key containing the messages (default: \"messages\")\n field_messages: str | None\n # Key containing the tools (default: \"tools\"). Must be a list[dict] and follow [JSON\n # schema](https://json-schema.org/learn/getting-started-step-by-step).\n field_tools: str | None\n\n message_field_role: str | None\n\n message_field_content: str | None\n # Mapping of properties from the input dataset to the chat template. (default:\n # message_property_mappings={'role':'role', 'content':'content'}) If a property exists\n # in the template but not in this mapping, the system will attempt to load it directly\n # from the message using the property name as the key. Example: In the mapping below,\n # 'from' is loaded from input dataset and used as 'role', while 'value' is loaded and\n # used as 'content' in the chat template.\n message_property_mappings: dict[str, str] | None\n # The key in the message turn that indicates via boolean whether tokens of a turn\n # should be considered for training. Useful to selectively train on certain turns\n # besides the `roles_to_train`.\n message_field_training: str | None\n # The key in the message turn that contains the training details. Useful to\n # selectively train on certain tokens in a turn. The value of the key is a List[Dict]\n # containing `begin_offset` (start character index in content), `end_offset` (end\n # character index in content), and `train` (boolean whether to train).\n message_field_training_detail: str | None\n # (for Qwen3 template only) Whether to split the assistant content based on a\n # reasoning trace inside delimited tags\n split_thinking: bool | None\n logprobs_field: str | None\n temperature: float | None\n # Roles to train on. The tokens from these roles will be considered for the loss.\n roles_to_train: list[str] | None\n # Which EOS tokens to train on in the conversation. Possible values are: all: train on\n # all EOS tokens, turn (default): train on the EOS token at the end of each trainable\n # turn, last: train on the last EOS token in the conversation\n train_on_eos: Literal['all', 'turn', 'last'] | None\n # Roles mapping in the messages. The format is {target_role: [source_roles]}. All\n # source roles will be mapped to the target role. The default is: user: [\"human\",\n # \"user\"], assistant: [\"gpt\", \"assistant\"], system: [\"system\"], tool: [\"tool\"]\n roles: dict[str, list[str]] | None\n # Whether to drop the system turn from the dataset. Only works with chat_template.\n # This does not drop the default system message from chat_template if it exists. If\n # you wish to, we recommend using a custom jinja template with the default system\n # message removed or adding a system turn with empty content.\n drop_system_message: bool | None\n # Trust remote code for untrusted source\n trust_remote_code: bool | None = False\n # The specific revision of the dataset to use when loading from the Hugging Face Hub.\n # This can be a commit hash, tag, or branch name. If not specified, the latest version\n # will be used. This parameter is ignored for local datasets.\n revision: str | None\n\n # For DPODataset:\n path: str | None\n split: str | None\n type: UserDefinedDPOType | str | None\n # For UserDefinedDPOType:\n field_system: str | None\n field_prompt: str | None\n field_chosen: str | None\n field_rejected: str | None\n prompt_format: str | None\n chosen_format: str | None\n rejected_format: str | None\n data_files: list[str] | None\n revision: str | None\n field_messages: str | None\n\n # For KTODataset:\n path: str | None\n split: str | None\n type: UserDefinedKTOType | str | None\n # For UserDefinedKTOType:\n field_system: str | None\n field_prompt: str | None\n field_completion: str | None\n field_label: bool | None\n prompt_format: str | None\n completion_format: str | None\n data_files: list[str] | None\n trust_remote_code: bool | None = False\n revision: str | None\n\n # For StepwiseSupervisedDataset:\n path: str | None\n split: str | None\n data_files: list[str] | None\n revision: str | None\n step_separator: str | None\n max_completion_length: int | None\n train_on_last_step_only: bool | None\n\n# A list of one or more datasets to eval the model with. You can use either\n# test_datasets, or val_set_size, but not both.\ntest_datasets: Annotated[list[SFTDataset | DPODataset | KTODataset | StepwiseSupervisedDataset], MinLen(1)] | None\n # For SFTDataset:\n # HuggingFace dataset repo | s3:// | gs:// | path to local file or directory\n path: str | None\n # name of dataset split to load from\n split: str | None\n # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]\n type: str | UserDefinedPrompterType | None\n # For UserDefinedPrompterType:\n # Custom user instruction prompt\n system_prompt: str | None\n # Use {system} as key to be replaced\n system_format: str | None\n field_system: str | None\n field_instruction: str | None\n field_input: str | None\n field_output: str | None\n\n # Customizable to be single line or multi-line. Use {instruction}/{input} as key to\n # be replaced. 'format' can include {input}\n format: str | None\n # 'no_input_format' cannot include {input}\n no_input_format: str | None\n input_transform: str | None\n # split dataset into N pieces (use with shards_idx)\n shards: int | None\n # the index of sharded dataset to use\n shards_idx: int | None\n # process dataset in N sequential chunks for memory efficiency (exclusive with\n # `shards`)\n preprocess_shards: int | None\n conversation: str | None\n\n # The name of the chat template to use for training, following values are supported:\n # tokenizer_default: Uses the chat template that is available in the\n # tokenizer_config.json. If the chat template is not available in the tokenizer, it\n # will raise an error. This is the default.\n # alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates\n # are available in the axolotl codebase at src/axolotl/utils/chat_templates.py.\n # tokenizer_default_fallback_*: where * is the name of the chat template to fallback\n # to if the tokenizer does not have a chat template else default to tokenizer. E.g.\n # tokenizer_default_fallback_chatml. jinja: Uses a custom jinja template for the chat\n # template. The custom jinja template should be provided in the chat_template_jinja\n # field.\n chat_template: ChatTemplate | str | None\n # Custom jinja chat template or path to jinja file. Used only if `chat_template:\n # jinja` or empty.\n chat_template_jinja: str | None\n # path to source data files\n data_files: str | list[str] | None\n input_format: str | None\n # name of dataset configuration to load\n name: str | None\n # defines the datatype when path is a file\n ds_type: str | None\n # For `completion` datasets only, uses the provided field instead of `text` column\n field: str | None\n field_human: str | None\n field_model: str | None\n # Key containing the messages (default: \"messages\")\n field_messages: str | None\n # Key containing the tools (default: \"tools\"). Must be a list[dict] and follow [JSON\n # schema](https://json-schema.org/learn/getting-started-step-by-step).\n field_tools: str | None\n\n message_field_role: str | None\n\n message_field_content: str | None\n # Mapping of properties from the input dataset to the chat template. (default:\n # message_property_mappings={'role':'role', 'content':'content'}) If a property exists\n # in the template but not in this mapping, the system will attempt to load it directly\n # from the message using the property name as the key. Example: In the mapping below,\n # 'from' is loaded from input dataset and used as 'role', while 'value' is loaded and\n # used as 'content' in the chat template.\n message_property_mappings: dict[str, str] | None\n # The key in the message turn that indicates via boolean whether tokens of a turn\n # should be considered for training. Useful to selectively train on certain turns\n # besides the `roles_to_train`.\n message_field_training: str | None\n # The key in the message turn that contains the training details. Useful to\n # selectively train on certain tokens in a turn. The value of the key is a List[Dict]\n # containing `begin_offset` (start character index in content), `end_offset` (end\n # character index in content), and `train` (boolean whether to train).\n message_field_training_detail: str | None\n # (for Qwen3 template only) Whether to split the assistant content based on a\n # reasoning trace inside delimited tags\n split_thinking: bool | None\n logprobs_field: str | None\n temperature: float | None\n # Roles to train on. The tokens from these roles will be considered for the loss.\n roles_to_train: list[str] | None\n # Which EOS tokens to train on in the conversation. Possible values are: all: train on\n # all EOS tokens, turn (default): train on the EOS token at the end of each trainable\n # turn, last: train on the last EOS token in the conversation\n train_on_eos: Literal['all', 'turn', 'last'] | None\n # Roles mapping in the messages. The format is {target_role: [source_roles]}. All\n # source roles will be mapped to the target role. The default is: user: [\"human\",\n # \"user\"], assistant: [\"gpt\", \"assistant\"], system: [\"system\"], tool: [\"tool\"]\n roles: dict[str, list[str]] | None\n # Whether to drop the system turn from the dataset. Only works with chat_template.\n # This does not drop the default system message from chat_template if it exists. If\n # you wish to, we recommend using a custom jinja template with the default system\n # message removed or adding a system turn with empty content.\n drop_system_message: bool | None\n # Trust remote code for untrusted source\n trust_remote_code: bool | None = False\n # The specific revision of the dataset to use when loading from the Hugging Face Hub.\n # This can be a commit hash, tag, or branch name. If not specified, the latest version\n # will be used. This parameter is ignored for local datasets.\n revision: str | None\n\n # For DPODataset:\n path: str | None\n split: str | None\n type: UserDefinedDPOType | str | None\n # For UserDefinedDPOType:\n field_system: str | None\n field_prompt: str | None\n field_chosen: str | None\n field_rejected: str | None\n prompt_format: str | None\n chosen_format: str | None\n rejected_format: str | None\n data_files: list[str] | None\n revision: str | None\n field_messages: str | None\n\n # For KTODataset:\n path: str | None\n split: str | None\n type: UserDefinedKTOType | str | None\n # For UserDefinedKTOType:\n field_system: str | None\n field_prompt: str | None\n field_completion: str | None\n field_label: bool | None\n prompt_format: str | None\n completion_format: str | None\n data_files: list[str] | None\n trust_remote_code: bool | None = False\n revision: str | None\n\n # For StepwiseSupervisedDataset:\n path: str | None\n split: str | None\n data_files: list[str] | None\n revision: str | None\n step_separator: str | None\n max_completion_length: int | None\n train_on_last_step_only: bool | None\n\n# If false, the datasets will not be shuffled and will keep their original order in\n# `datasets`. The same applies to the `test_datasets` option and the\n# `pretraining_dataset` option. Default is true.\nshuffle_merged_datasets: bool | None = True\n# If true, each dataset in `datasets` will be shuffled before merging. This allows\n# curriculum learning strategies to be applied at the dataset level. Default is false.\nshuffle_before_merging_datasets: bool | None = False\n# Axolotl attempts to save the dataset as an arrow after packing the data together so\n# subsequent training attempts load faster, relative path\ndataset_prepared_path: str | None\n# Num shards for whole dataset\ndataset_shard_num: int | None\n# Index of shard to use for whole dataset\ndataset_shard_idx: int | None\nskip_prepare_dataset: bool | None = False\n# Number of shards to save the prepared dataset\nnum_dataset_shards_to_save: int | None\n\n# Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize\npretraining_dataset: Annotated[list[PretrainingDataset | SFTDataset], MinLen(1)] | None\n # For PretrainingDataset:\n name: str | None\n path: str | None\n split: str | None = train\n text_column: str | None = text\n type: str | None = pretrain\n trust_remote_code: bool | None = False\n data_files: str | None\n skip: int | None\n\n # For SFTDataset:\n # HuggingFace dataset repo | s3:// | gs:// | path to local file or directory\n path: str | None\n # name of dataset split to load from\n split: str | None\n # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]\n type: str | UserDefinedPrompterType | None\n # For UserDefinedPrompterType:\n # Custom user instruction prompt\n system_prompt: str | None\n # Use {system} as key to be replaced\n system_format: str | None\n field_system: str | None\n field_instruction: str | None\n field_input: str | None\n field_output: str | None\n\n # Customizable to be single line or multi-line. Use {instruction}/{input} as key to\n # be replaced. 'format' can include {input}\n format: str | None\n # 'no_input_format' cannot include {input}\n no_input_format: str | None\n input_transform: str | None\n # split dataset into N pieces (use with shards_idx)\n shards: int | None\n # the index of sharded dataset to use\n shards_idx: int | None\n # process dataset in N sequential chunks for memory efficiency (exclusive with\n # `shards`)\n preprocess_shards: int | None\n conversation: str | None\n\n # The name of the chat template to use for training, following values are supported:\n # tokenizer_default: Uses the chat template that is available in the\n # tokenizer_config.json. If the chat template is not available in the tokenizer, it\n # will raise an error. This is the default.\n # alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates\n # are available in the axolotl codebase at src/axolotl/utils/chat_templates.py.\n # tokenizer_default_fallback_*: where * is the name of the chat template to fallback\n # to if the tokenizer does not have a chat template else default to tokenizer. E.g.\n # tokenizer_default_fallback_chatml. jinja: Uses a custom jinja template for the chat\n # template. The custom jinja template should be provided in the chat_template_jinja\n # field.\n chat_template: ChatTemplate | str | None\n # Custom jinja chat template or path to jinja file. Used only if `chat_template:\n # jinja` or empty.\n chat_template_jinja: str | None\n # path to source data files\n data_files: str | list[str] | None\n input_format: str | None\n # name of dataset configuration to load\n name: str | None\n # defines the datatype when path is a file\n ds_type: str | None\n # For `completion` datasets only, uses the provided field instead of `text` column\n field: str | None\n field_human: str | None\n field_model: str | None\n # Key containing the messages (default: \"messages\")\n field_messages: str | None\n # Key containing the tools (default: \"tools\"). Must be a list[dict] and follow [JSON\n # schema](https://json-schema.org/learn/getting-started-step-by-step).\n field_tools: str | None\n\n message_field_role: str | None\n\n message_field_content: str | None\n # Mapping of properties from the input dataset to the chat template. (default:\n # message_property_mappings={'role':'role', 'content':'content'}) If a property exists\n # in the template but not in this mapping, the system will attempt to load it directly\n # from the message using the property name as the key. Example: In the mapping below,\n # 'from' is loaded from input dataset and used as 'role', while 'value' is loaded and\n # used as 'content' in the chat template.\n message_property_mappings: dict[str, str] | None\n # The key in the message turn that indicates via boolean whether tokens of a turn\n # should be considered for training. Useful to selectively train on certain turns\n # besides the `roles_to_train`.\n message_field_training: str | None\n # The key in the message turn that contains the training details. Useful to\n # selectively train on certain tokens in a turn. The value of the key is a List[Dict]\n # containing `begin_offset` (start character index in content), `end_offset` (end\n # character index in content), and `train` (boolean whether to train).\n message_field_training_detail: str | None\n # (for Qwen3 template only) Whether to split the assistant content based on a\n # reasoning trace inside delimited tags\n split_thinking: bool | None\n logprobs_field: str | None\n temperature: float | None\n # Roles to train on. The tokens from these roles will be considered for the loss.\n roles_to_train: list[str] | None\n # Which EOS tokens to train on in the conversation. Possible values are: all: train on\n # all EOS tokens, turn (default): train on the EOS token at the end of each trainable\n # turn, last: train on the last EOS token in the conversation\n train_on_eos: Literal['all', 'turn', 'last'] | None\n # Roles mapping in the messages. The format is {target_role: [source_roles]}. All\n # source roles will be mapped to the target role. The default is: user: [\"human\",\n # \"user\"], assistant: [\"gpt\", \"assistant\"], system: [\"system\"], tool: [\"tool\"]\n roles: dict[str, list[str]] | None\n # Whether to drop the system turn from the dataset. Only works with chat_template.\n # This does not drop the default system message from chat_template if it exists. If\n # you wish to, we recommend using a custom jinja template with the default system\n # message removed or adding a system turn with empty content.\n drop_system_message: bool | None\n # Trust remote code for untrusted source\n trust_remote_code: bool | None = False\n # The specific revision of the dataset to use when loading from the Hugging Face Hub.\n # This can be a commit hash, tag, or branch name. If not specified, the latest version\n # will be used. This parameter is ignored for local datasets.\n revision: str | None\n\n# The maximum number of processes to use while preprocessing your input dataset. This\n# defaults to `os.cpu_count()` if not set. For Runpod VMs, it will default to number of\n# vCPUs via RUNPOD_CPU_COUNT.\ndataset_processes: int | None\n# Deduplicates datasets and test_datasets with identical entries\ndataset_exact_deduplication: bool | None\n# Keep dataset in memory while preprocessing. Only needed if cached dataset is taking\n# too much storage\ndataset_keep_in_memory: bool | None\ndataloader_pin_memory: bool | None\ndataloader_num_workers: int | None\ndataloader_prefetch_factor: int | None\ndataloader_drop_last: bool | None\n\naccelerator_config: dict[str, Any] | None\n\nremove_unused_columns: bool | None\n\n# Push prepared dataset to hub - repo_org/repo_name\npush_dataset_to_hub: str | None\n# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private\n# datasets. Required to be true when used in combination with `push_dataset_to_hub`\nhf_use_auth_token: bool | None\n\ndevice: Any | None\n# Passed through to transformers when loading the model when launched without\n# accelerate. Use `sequential` when training w/ model parallelism to limit memory\ndevice_map: Any | None\nworld_size: int | None\n# Don't mess with this, it's here for accelerate and torchrun\nlocal_rank: int | None\nddp: bool | None\n\n# Seed for reproducibility\nseed: int | None\n# Advanced DDP Arguments - timeout\nddp_timeout: int | None\n# Advanced DDP Arguments - bucket cap in MB\nddp_bucket_cap_mb: int | None\n# Advanced DDP Arguments - broadcast buffers\nddp_broadcast_buffers: bool | None\nddp_find_unused_parameters: bool | None\n\n# Approximate number of predictions sent to wandb depending on batch size. Enabled above\n# 0. Default is 0\neval_table_size: int | None\n# Total number of tokens generated for predictions sent to wandb. Default is 128\neval_max_new_tokens: int | None\n# Whether to run causal language model evaluation for metrics in\n# `eval_causal_lm_metrics`\ndo_causal_lm_eval: bool | None\n# HF evaluate metrics used during evaluation. Default is ['sacrebleu', 'comet', 'ter',\n# 'chrf', 'perplexity']\neval_causal_lm_metrics: list[str] | None\ndo_bench_eval: bool | None\nbench_dataset: str | None\nbench_split: str | None\nmetric_for_best_model: str | None\ngreater_is_better: bool | None\n\n# High loss value, indicating the learning has broken down (a good estimate is ~2 times\n# the loss at the start of training)\nloss_watchdog_threshold: float | None\n# Number of high-loss steps in a row before the trainer aborts (default: 3)\nloss_watchdog_patience: int | None\n\n# Run garbage collection every `gc_steps` steps. -1 will run on epoch end and before\n# evaluations. Default is 0 (disabled).\ngc_steps: int | None\n\n# Use CUDA bf16. bool or 'full' for `bf16_full_eval`, or 'auto' for automatic detection.\n# require >=ampere\nbf16: Literal['auto'] | bool | None = auto\n# Use CUDA fp16\nfp16: bool | None\n# Enable FP8 mixed precision training using TorchAO. Best used in combination with\n# torch.compile.\nfp8: bool | None\n# Enable FSDP float8 all-gather optimization for FP8 training. Can improve training\n# speed by 10-15% when FSDP is enabled.\nfp8_enable_fsdp_float8_all_gather: bool | None\n# No AMP (automatic mixed precision) - require >=ampere\nbfloat16: bool | None\n# No AMP (automatic mixed precision)\nfloat16: bool | None\n# Use CUDA tf32 - require >=ampere\ntf32: bool | None\nfloat32: bool | None\n\n# Whether to use gradient checkpointing. Available options are: true, false, 'offload',\n# 'offload_disk'.\n# https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing\ngradient_checkpointing: Literal['offload', 'offload_disk'] | bool | None = False\n# Additional kwargs to pass to the trainer for gradient checkpointing\ngradient_checkpointing_kwargs: dict[str, Any] | None\n# Whether to offload activations. Available options are: true, false, 'legacy', 'disk'.\nactivation_offloading: Literal['legacy', 'disk'] | bool | None = False\n\nunfrozen_parameters: list[str] | None\n\n# The maximum length of an input to train with, this should typically be less than 2048\n# as most models have a token/context limit of 2048\nsequence_len: int = 512\n# The maximum length of an input for evaluation. If not specified, defaults to\n# sequence_len\neval_sequence_len: int | None\nmin_sample_len: int | None\n# maximum prompt length for RL training\nmax_prompt_len: int = 512\n# Use efficient multi-packing with block diagonal attention and per sequence\n# position_ids. Recommend set to 'true'\nsample_packing: bool | None\n# The number of samples packed at a time. Increasing the following values helps with\n# packing, but usually only slightly (<%1.)\nsample_packing_group_size: int | None = 100000\n# The number of samples which can be packed into one sequence. Increase if using a large\n# sequence_len with many short samples.\nsample_packing_bin_size: int | None = 200\n# Whether to pack samples sequentially\nsample_packing_sequentially: bool | None\n# The multiprocessing start method to use for packing. Should be 'fork', 'spawn' or\n# 'forkserver'\nsample_packing_mp_start_method: str | None\n# Set to 'false' if getting errors during eval with sample_packing on\neval_sample_packing: bool | None\n# Pad inputs so each step uses constant sized buffers. This will reduce memory\n# fragmentation and may prevent OOMs, by re-using memory more efficiently. Defaults to\n# True if `sample_packing` enabled\npad_to_sequence_len: bool | None\n# Whether to use sequential sampling for curriculum learning\ncurriculum_sampling: bool | None\nmultipack_real_batches: bool | None\n# whether to concatenate samples during pretraining\npretraining_sample_concatenation: bool | None\n\n# Use batch flattening for speedups when not using sample_packing\nbatch_flattening: Literal['auto'] | bool | None\n\nuse_pose: bool | None\npose_split_on_token_ids: list[int] | None\npose_max_context_len: int | None\npose_num_chunks: int | None\n\npretrain_multipack_buffer_size: int | None = 10000\n# whether to prevent cross attention for packed sequences during pretraining\npretrain_multipack_attn: bool | None = True\n\n# Whether to use xformers attention patch https://github.com/facebookresearch/xformers\nxformers_attention: bool | None\n# Whether to use scaled-dot-product attention https://pytorch.org/docs/stable/generated/\n# torch.nn.functional.scaled_dot_product_attention.html\nsdp_attention: bool | None\n# Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf\ns2_attention: bool | None\nflex_attention: bool | None\nflex_attn_compile_kwargs: dict[str, Any] | None\n# Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention\nflash_attention: bool | None\n# Whether to use flash-attention cross entropy implementation - advanced use only\nflash_attn_cross_entropy: bool | None\n# Whether to use flash-attention rms norm implementation - advanced use only\nflash_attn_rms_norm: bool | None\n# Whether to fuse part of the MLP into a single operation\nflash_attn_fuse_mlp: bool | None\n# Whether to use bettertransformers\nflash_optimum: bool | None\n\neager_attention: bool | None\n\n# Specify a custom attention implementation, used mostly for kernels.\nattn_implementation: str | None\n\nunsloth_cross_entropy_loss: bool | None\nunsloth_lora_mlp: bool | None\nunsloth_lora_qkv: bool | None\nunsloth_lora_o: bool | None\nunsloth_rms_norm: bool | None\nunsloth_rope: bool | None\n\n# Apply custom LoRA autograd functions and activation function Triton kernels for speed\n# and memory savings. See: https://docs.axolotl.ai/docs/lora_optims.html\nlora_mlp_kernel: bool | None\n# Apply custom LoRA autograd functions and activation function Triton kernels for speed\n# and memory savings. See: https://docs.axolotl.ai/docs/lora_optims.html\nlora_qkv_kernel: bool | None\n# Apply custom LoRA autograd functions and activation function Triton kernels for speed\n# and memory savings. See: https://docs.axolotl.ai/docs/lora_optims.html\nlora_o_kernel: bool | None\n\n# Whether to use chunked cross entropy loss for memory efficiency\nchunked_cross_entropy: bool | None\n# Number of chunks to use for chunked cross entropy loss\nchunked_cross_entropy_num_chunks: int | None\n\n# Whether to use ALST tiled mlp for memory efficient long context\ntiled_mlp: bool | None\n\n# Number of shards to use for ALST tiled mlp. If unset, it will be set based on\n# seqlen/hidden_size\ntiled_mlp_num_shards: int | None\n\n# Whether to use original mlp for ALST tiled mlp. Otherwise uses a generic MLP based on\n# llama.\ntiled_mlp_use_original_mlp: bool | None = True\n\nllama4_linearized_experts: bool | None\n\n# Deepspeed config path. e.g., deepspeed_configs/zero3.json\ndeepspeed: str | dict[str, Any] | None\n# Whether to use deepcompile for faster training with deepspeed\ndeepcompile: bool | None\n# FSDP configuration\nfsdp: list[str] | None\n\n# FSDP configuration options\nfsdp_config: dict[str, Any] | None\n# FSDP version\nfsdp_version: int | None\nfsdp_final_state_dict_type: Literal['FULL_STATE_DICT', 'LOCAL_STATE_DICT', 'SHARDED_STATE_DICT'] | None\n\n# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for\n# no eval.\nval_set_size: float | None = 0.0\n\n# Number of devices to shard across. If not set, will use all available devices.\ndp_shard_size: int | None\n# Number of devices to replicate across.\ndp_replicate_size: int | None\n# Deprecated: use `context_parallel_size` instead\nsequence_parallel_degree: int | None\n# Set to a divisor of the number of GPUs available to split sequences into chunks of\n# equal size. Use in long context training to prevent OOM when sequences cannot fit into\n# a single GPU's VRAM. E.g., if 4 GPUs are available, set this value to 2 to split each\n# sequence into two equal-sized subsequences, or set to 4 to split into four equal-sized\n# subsequences. See https://docs.axolotl.ai/docs/sequence_parallelism.html for more\n# details.\ncontext_parallel_size: int | None\n# Optional; strides across the key dimension. Larger values use more memory but should\n# make training faster. Must evenly divide the number of KV heads in your model.\nheads_k_stride: int | None\n# One of 'varlen_llama3', 'batch_ring', 'batch_zigzag', 'batch_stripe'. Defaults to\n# 'varlen_llama3' in the sample packing case, and 'batch_ring' in the non-sample packing\n# case.\nring_attn_func: RingAttnFunc | None\n# Number of tensor parallel processes in TP group. Only supported with DeepSpeed AutoTP.\ntensor_parallel_size: int | None\n\n# Add or change special tokens. If you add tokens here, you don't need to add them to\n# the `tokens` list.\nspecial_tokens: SpecialTokensConfig | None\n # For SpecialTokensConfig:\n bos_token: str | None\n eos_token: str | None\n pad_token: str | None\n unk_token: str | None\n additional_special_tokens: list[str] | None\n\n# Add extra tokens to the tokenizer\ntokens: list[str] | None\n# Mapping token_id to new_token_string to override reserved added_tokens in the\n# tokenizer. Only works for tokens that are not part of the base vocab (aka are\n# added_tokens). Can be checked if they exist in tokenizer.json added_tokens.\nadded_tokens_overrides: dict[int, str] | None\n\n# Whether to use torch.compile and which backend to use. setting to `auto` will enable\n# torch compile when torch>=2.6.0\ntorch_compile: Literal['auto'] | bool | None\n# Backend to use for torch.compile\ntorch_compile_backend: str | None\ntorch_compile_mode: Literal['default', 'reduce-overhead', 'max-autotune'] | None\n\n# Maximum number of iterations to train for. It precedes num_epochs which means that if\n# both are set, num_epochs will not be guaranteed. e.g., when 1 epoch is 1000 steps =>\n# `num_epochs: 2` and `max_steps: 100` will train for 100 steps\nmax_steps: int | None\n# Number of warmup steps. Cannot use with warmup_ratio\nwarmup_steps: int | None\n# Warmup ratio. Cannot use with warmup_steps\nwarmup_ratio: float | None\n# Leave empty to eval at each epoch, integer for every N steps. float for fraction of\n# total steps\neval_steps: int | float | None\n# Number of times per epoch to run evals, mutually exclusive with eval_steps\nevals_per_epoch: int | None\n# Set to `no` to skip evaluation, `epoch` at end of each epoch, leave empty to infer\n# from `eval_steps`\neval_strategy: str | None\n\n# Leave empty to save at each epoch, integer for every N steps. float for fraction of\n# total steps\nsave_steps: int | float | None\n# Number of times per epoch to save a checkpoint, mutually exclusive with save_steps\nsaves_per_epoch: int | None\n# Set to `no` to skip checkpoint saves, `epoch` at end of each epoch, `best` when better\n# result is achieved, leave empty to infer from `save_steps`\nsave_strategy: str | None\n# Checkpoints saved at a time\nsave_total_limit: int | None\n# Whether to checkpoint a model after the first step of training. Defaults to False.\nsave_first_step: bool | None\n\n# Logging frequency\nlogging_steps: int | None\n# Stop training after this many evaluation losses have increased in a row. https://huggi\n# ngface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppin\n# gCallback\nearly_stopping_patience: int | None\nload_best_model_at_end: bool | None = False\n# Save only the model weights, skipping the optimizer. Using this means you can't resume\n# from checkpoints.\nsave_only_model: bool | None = False\n# Use tensorboard for logging\nuse_tensorboard: bool | None\n# Enable the pytorch profiler to capture the first N steps of training to the\n# output_dir. see https://pytorch.org/blog/understanding-gpu-memory-1/ for more\n# information. Snapshots can be visualized @ https://pytorch.org/memory_viz\nprofiler_steps: int | None\n# Which step to start the profiler at. Useful for only capturing a few steps mid-run.\nprofiler_steps_start: int | None = 0\n# bool of whether to include tokens trainer per second in the training metrics. This\n# iterates over the entire dataset once, so it takes some time.\ninclude_tokens_per_second: bool | None\n\n# NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to\n# add noise to embeddings. Currently only supported on Llama and Mistral\nneftune_noise_alpha: float | None\n\n# Parameter controlling the relative ratio loss weight in the ORPO loss. Passed to\n# `beta` in `ORPOConfig` due to trl mapping.\norpo_alpha: float | None\n# Weighting of NLL term in loss from RPO paper\nrpo_alpha: float | None\n# Target reward margin for the SimPO loss\nsimpo_gamma: float | None\n# Weight of the BC regularizer\ncpo_alpha: float | None\n\n# Factor for desirable loss term in KTO loss\nkto_desirable_weight: float | None\n# Factor for undesirable loss term in KTO loss\nkto_undesirable_weight: float | None\n# The beta parameter for the RL training\nrl_beta: float | None\n\n# Defines the max memory usage per gpu on the system. Passed through to transformers\n# when loading the model.\nmax_memory: dict[int | Literal['cpu', 'disk'], int | str] | None\n# Limit the memory for all available GPUs to this amount (if an integer, expressed in\n# gigabytes); default: unset\ngpu_memory_limit: int | str | None\n# Whether to use low_cpu_mem_usage\nlow_cpu_mem_usage: bool | None\n\n# The name of the chat template to use for training, following values are supported:\n# tokenizer_default: Uses the chat template that is available in the\n# tokenizer_config.json. If the chat template is not available in the tokenizer, it will\n# raise an error. This is the default value.\n# alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates\n# are available in the axolotl codebase at src/axolotl/utils/chat_templates.py.\n# tokenizer_default_fallback_*: where * is the name of the chat template to fallback to.\n# E.g. tokenizer_default_fallback_chatml. This is useful when the chat template is not\n# available in the tokenizer. jinja: Uses a custom jinja template for the chat template.\n# The custom jinja template should be provided in the chat_template_jinja field. The\n# selected chat template will be saved to the tokenizer_config.json for easier\n# inferencing\nchat_template: ChatTemplate | Annotated[str, StringConstraints(pattern='^tokenizer_default_fallback_')] | None\n# Custom jinja template or path to jinja file for chat template. This will be only used\n# if chat_template is set to `jinja` or `null` (in which case chat_template is\n# automatically set to `jinja`). Default is null.\nchat_template_jinja: str | None\n# Additional kwargs to pass to the chat template. This is useful for customizing the\n# chat template. For example, you can pass `thinking=False` to add a generation prompt\n# to the chat template.\nchat_template_kwargs: dict[str, Any] | None\n# Custom EOT (End-of-Turn) tokens to mask/unmask during training. These tokens mark the\n# boundaries between conversation turns. For example: ['/INST', '</s>',\n# '[/SYSTEM_PROMPT]']. If not specified, defaults to just the model's eos_token. This is\n# useful for templates that use multiple delimiter tokens.\neot_tokens: list[str] | None\n# Changes the default system message. Currently only supports chatml.\ndefault_system_message: str | None\n\nfix_untrained_tokens: int | list[int] | None\n\nis_preprocess: bool | None\npreprocess_iterable: bool | None\n\n# Total number of tokens - internal use\ntotal_num_tokens: int | None\ntotal_supervised_tokens: int | None\n# You can set these packing optimizations AFTER starting a training at least once. The\n# trainer will provide recommended values for these values.\nsample_packing_eff_est: float | None\naxolotl_config_path: str | None\n\n# Internal use only - Used to identify which the model is based on\nis_falcon_derived_model: bool | None\n# Internal use only - Used to identify which the model is based on\nis_llama_derived_model: bool | None\n# Internal use only - Used to identify which the model is based on. Please note that if\n# you set this to true, `padding_side` will be set to 'left' by default\nis_mistral_derived_model: bool | None\n# Internal use only - Used to identify which the model is based on\nis_qwen_derived_model: bool | None\n\n# Add plugins to extend the pipeline. See `src/axolotl/integrations` for the available\n# plugins or doc below for more details.\n# https://docs.axolotl.ai/docs/custom_integrations.html\nplugins: list[str] | None\n\n# This is the huggingface model that contains *.pt, *.safetensors, or *.bin files. This\n# can also be a relative path to a model on disk\nbase_model: str (required)\n# If the base_model repo on hf hub doesn't include configuration .json files, You can\n# set that here, or leave this empty to default to base_model\nbase_model_config: str | None\ncls_model_config: str | None\n# Optional tokenizer configuration path in case you want to use a different tokenizer\n# than the one defined in the base model\ntokenizer_config: str | None\n# use_fast option for tokenizer loading from_pretrained, default to True\ntokenizer_use_fast: bool | None\n# Whether to use the legacy tokenizer setting, defaults to True\ntokenizer_legacy: bool | None\n# Whether to use mistral-common tokenizer. If set to True, it will use the mistral-\n# common tokenizer.\ntokenizer_use_mistral_common: bool | None\n# Corresponding tokenizer for the model AutoTokenizer is a good choice\ntokenizer_type: str | None\n# transformers processor class\nprocessor_type: str | None\n# Trust remote code for untrusted source\ntrust_remote_code: bool | None\n\n# Don't move the model to the device before sharding. This is an experimental feature\n# that may be included in the future as the default.\nexperimental_skip_move_to_device: bool | None\n\n# Use custom kernels, e.g. MegaBlocks.\nuse_kernels: bool | None\n\n# Model loading quantization config\nmodel_quantization_config: Literal['Mxfp4Config'] | None\n# kwargs for model quantization config\nmodel_quantization_config_kwargs: dict[str, Any] | None\n\n# Where to save the full-finetuned model to\noutput_dir: str = ./model-out\n# push checkpoints to hub\nhub_model_id: str | None\n# how to push checkpoints to hub\nhub_strategy: str | None\n# Save model as safetensors (require safetensors package). Default True\nsave_safetensors: bool | None = True\n\n# This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer\nload_in_8bit: bool | None = False\n# Use bitsandbytes 4 bit\nload_in_4bit: bool | None = False\n\n# If you want to use 'lora' or 'qlora' or leave blank to train all parameters in\n# original model\nadapter: str | None\n# If you already have a lora model trained that you want to load, put that here. This\n# means after training, if you want to test the model, you should set this to the value\n# of `output_dir`. Note that if you merge an adapter to the base model, a new\n# subdirectory `merged` will be created under the `output_dir`.\nlora_model_dir: str | None\nlora_r: int | None\nlora_alpha: int | None\nlora_fan_in_fan_out: bool | None\nlora_target_modules: str | list[str] | None\nlora_target_parameters: str | list[str] | None\n# If true, will target all linear modules\nlora_target_linear: bool | None\n# If you added new tokens to the tokenizer, you may need to save some LoRA modules\n# because they need to know the new tokens. For LLaMA and Mistral, you need to save\n# `embed_tokens` and `lm_head`. It may vary for other models. `embed_tokens` converts\n# tokens to embeddings, and `lm_head` converts embeddings to token probabilities.\nlora_modules_to_save: list[str] | None\nlora_dropout: float | None = 0.0\n# The layer indices to transform, otherwise, apply to all layers\npeft_layers_to_transform: list[int] | None\npeft_layers_pattern: list[str] | None\n\npeft: PeftConfig | None\n # For PeftConfig:\n # Configuration options for loftq initialization for LoRA\n loftq_config: LoftQConfig | None\n # For LoftQConfig:\n # typically 4 bits\n loftq_bits: int = 4\n\n# Whether to use DoRA.\npeft_use_dora: bool | None\n# Whether to use RSLoRA.\npeft_use_rslora: bool | None\n# List of layer indices to replicate.\npeft_layer_replication: list[tuple[int, int]] | None\n# How to initialize LoRA weights. Default to True which is MS original implementation.\npeft_init_lora_weights: bool | str | None\n\n# load qlora model in sharded format for FSDP using answer.ai technique.\nqlora_sharded_model_loading: bool | None = False\n# Do the LoRA/PEFT loading on CPU -- this is required if the base model is so large it\n# takes up most or all of the available GPU VRAM, e.g. during a model and LoRA merge\nlora_on_cpu: bool | None\n# Whether you are training a 4-bit GPTQ quantized model\ngptq: bool | None\n# optional overrides to the bnb 4bit quantization configuration\nbnb_config_kwargs: dict[str, Any] | None\n\n# loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4.\nloraplus_lr_ratio: float | None\n# loraplus learning rate for lora embedding layers. Default value is 1e-6.\nloraplus_lr_embedding: float | None = 1e-06\n\nmerge_lora: bool | None\n\n# Whether to use ReLoRA. Use with jagged_restart_*steps options.\nrelora: bool | None\n# threshold for optimizer magnitude when pruning\nrelora_prune_ratio: float | None\n# True to perform lora weight merges on cpu during restarts, for modest gpu memory\n# savings\nrelora_cpu_offload: bool | None\n\n# how often to reset for jagged restarts\njagged_restart_steps: int | None\n# how many warmup steps to take after reset for jagged restarts\njagged_restart_warmup_steps: int | None\n# how many anneal steps to take before reset for jagged restarts\njagged_restart_anneal_steps: int | None\n\n# If greater than 1, backpropagation will be skipped and the gradients will be\n# accumulated for the given number of steps.\ngradient_accumulation_steps: int | None = 1\n# The number of samples to include in each batch. This is the number of samples sent to\n# each GPU. Batch size per gpu = micro_batch_size * gradient_accumulation_steps\nmicro_batch_size: int | None = 1\n# Total batch size, we do not recommended setting this manually\nbatch_size: int | None\n# per gpu micro batch size for evals, defaults to value of micro_batch_size\neval_batch_size: int | None\n\n# whether to find batch size that fits in memory. Passed to underlying transformers\n# Trainer\nauto_find_batch_size: bool | None\n\n# Whether to mask out or include the human's prompt from the training labels\ntrain_on_inputs: bool | None = False\n# Group similarly sized data to minimize padding. May be slower to start, as it must\n# download and sort the entire dataset. Note that training loss may have an oscillating\n# pattern with this enabled.\ngroup_by_length: bool | None\n\nlearning_rate: str | float (required)\nembedding_lr: float | None\nembedding_lr_scale: float | None\n# Specify weight decay\nweight_decay: float | None = 0.0\n# Specify optimizer\noptimizer: OptimizerNames | CustomSupportedOptimizers | None = OptimizerNames.ADAMW_TORCH_FUSED\n# Dictionary of arguments to pass to the optimizer\noptim_args: str | dict[str, Any] | None\n# The target modules to optimize, i.e. the module names that you would like to train,\n# right now this is used only for GaLore algorithm\noptim_target_modules: list[str] | Literal['all_linear'] | None\n# Path to torch distx for optim 'adamw_anyprecision'\ntorchdistx_path: str | None\nlr_scheduler: SchedulerType | Literal['one_cycle'] | Literal['rex'] | None = SchedulerType.COSINE\n# Specify a scheduler and kwargs to use with the optimizer\nlr_scheduler_kwargs: dict[str, Any] | None\nlr_quadratic_warmup: bool | None\n# decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of\n# peak lr\ncosine_min_lr_ratio: float | None\n# freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means\n# start cosine_min_lr at 80% of training step\ncosine_constant_lr_ratio: float | None\n# Learning rate div factor\nlr_div_factor: float | None\n\nlr_groups: list[LrGroup] | None\n # For LrGroup:\n name: str (required)\n modules: list[str] (required)\n lr: float (required)\n\n# adamw hyperparams\nadam_epsilon: float | None\n# only used for CAME Optimizer\nadam_epsilon2: float | None\n# adamw hyperparams\nadam_beta1: float | None\n# adamw hyperparams\nadam_beta2: float | None\n# only used for CAME Optimizer\nadam_beta3: float | None\n\n# Dion Optimizer learning rate\ndion_lr: float | None\n# Dion Optimizer momentum\ndion_momentum: float | None\n# Dion Optimizer: r/d fraction for low-rank approximation. Used to compute the low-rank\n# dimension.\ndion_rank_fraction: float | None = 1.0\n# Dion Optimizer: Round up the low-rank dimension to a multiple of this number. This may\n# be useful to ensure even sharding.\ndion_rank_multiple_of: int | None = 1\n\n# Gradient clipping max norm\nmax_grad_norm: float | None\nnum_epochs: float = 1.0\n\nuse_wandb: bool | None\n# Set the name of your wandb run\nwandb_name: str | None\n# Set the ID of your wandb run\nwandb_run_id: str | None\n# \"offline\" to save run metadata locally and not sync to the server, \"disabled\" to turn\n# off wandb\nwandb_mode: str | None\n# Your wandb project name\nwandb_project: str | None\n# A wandb Team name if using a Team\nwandb_entity: str | None\nwandb_watch: str | None\n# \"checkpoint\" to log model to wandb Artifacts every `save_steps` or \"end\" to log only\n# at the end of training\nwandb_log_model: str | None\n\nuse_mlflow: bool | None\n# URI to mlflow\nmlflow_tracking_uri: str | None\n# Your experiment name\nmlflow_experiment_name: str | None\n# Your run name\nmlflow_run_name: str | None\n# set to true to copy each saved checkpoint on each save to mlflow artifact registry\nhf_mlflow_log_artifacts: bool | None\n\n# Enable or disable Comet integration.\nuse_comet: bool | None\n# API key for Comet. Recommended to set via `comet login`.\ncomet_api_key: str | None\n# Workspace name in Comet. Defaults to the user's default workspace.\ncomet_workspace: str | None\n# Project name in Comet. Defaults to Uncategorized.\ncomet_project_name: str | None\n# Identifier for the experiment. Used to append data to an existing experiment or\n# control the key of new experiments. Default to a random key.\ncomet_experiment_key: str | None\n# Create a new experiment (\"create\") or log to an existing one (\"get\"). Default\n# (\"get_or_create\") auto-selects based on configuration.\ncomet_mode: str | None\n# Set to True to log data to Comet server, or False for offline storage. Default is\n# True.\ncomet_online: bool | None\n# Dictionary for additional configuration settings, see the doc for more details.\ncomet_experiment_config: dict[str, Any] | None\n\n# the number of activate layers in LISA\nlisa_n_layers: int | None\n# how often to switch layers in LISA\nlisa_step_interval: int | None\n# path under the model to access the layers\nlisa_layers_attribute: str | None = model.layers\n\ngradio_title: str | None\ngradio_share: bool | None\ngradio_server_name: str | None\ngradio_server_port: int | None\ngradio_max_new_tokens: int | None\ngradio_temperature: float | None\n\nuse_ray: bool = False\nray_run_name: str | None\nray_num_workers: int = 1\nresources_per_worker: dict\n\n# The size of the image to resize to. It can be an integer (resized into padded-square\n# image) or a tuple (width, height).If not provided, we will attempt to load from\n# preprocessor.size, otherwise, images won't be resized.\nimage_size: int | tuple[int, int] | None\n# The resampling algorithm to use for image resizing. Default is bilinear. Please refer\n# to PIL.Image.Resampling for more details.\nimage_resize_algorithm: Literal['bilinear', 'bicubic', 'lanczos'] | Resampling | None\n\n# optional overrides to the base model configuration\noverrides_of_model_config: dict[str, Any] | None\n# optional overrides the base model loading from_pretrained\noverrides_of_model_kwargs: dict[str, Any] | None\n# If you want to specify the type of model to load, AutoModelForCausalLM is a good\n# choice too\ntype_of_model: str | None\n# You can specify to choose a specific model revision from huggingface hub\nrevision_of_model: str | None\n\nmax_packed_sequence_len: int | None\nrope_scaling: Any | None\nnoisy_embedding_alpha: float | None\ndpo_beta: float | None\nevaluation_strategy: str | None", + "text": "# Allow overwrite yml config using from cli\nstrict: bool | None = False\n# Resume from a specific checkpoint dir\nresume_from_checkpoint: str | None\n# If resume_from_checkpoint isn't set and you simply want it to start where it left off.\n# Be careful with this being turned on between different models.\nauto_resume_from_checkpoints: bool | None\n# Resize the model embeddings when new tokens are added to multiples of 32. This is\n# reported to improve training speed on some models\nresize_token_embeddings_to_32x: bool | None\nmean_resizing_embeddings: bool | None = False\n\n# Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.\nshrink_embeddings: bool | None\n# Don't upcast the embeddings to float32 when using PEFT. Useful for low-VRAM GPUs\nembeddings_skip_upcast: bool | None\n\n# module to custom trainer class to use for training\ntrainer_cls: str | None\n\n# Use RL training: 'dpo', 'ipo', 'kto', 'simpo', 'orpo', 'grpo'\nrl: RLType | None\n\ntrl: TRLConfig | None\n # For TRLConfig:\n # Beta parameter for the RL training. Same as `rl_beta`. Use\n beta: float | None\n # Maximum length of the completion for RL training.\n max_completion_length: int | None\n\n # Whether to use VLLM for RL training.\n use_vllm: bool = False\n # VLLM mode to use, one of 'server' or 'colocate'\n vllm_mode: Literal['server', 'colocate'] | None\n # Host of the vLLM server to connect to.\n vllm_server_host: str | None = 0.0.0.0\n # Port of the vLLM server to connect to.\n vllm_server_port: int | None = 8000\n # Total timeout (in seconds) to wait for the vLLM server to respond.\n vllm_server_timeout: int | None\n # Regex for vLLM guided decoding.\n vllm_guided_decoding_regex: str | None\n\n # List of reward functions to load. Paths must be importable from current dir.\n reward_funcs: list[str] | None\n # List of reward weights for the reward functions.\n reward_weights: list[float] | None\n # Number of generations to sample.\n num_generations: int | None\n # Whether to log completions.\n log_completions: bool | None = False\n # Number of completions to print when log_completions is True.\n num_completions_to_print: int | None\n # Controls whether importance sampling ratios are computed at the `'token'` or\n # `'sequence'` level. For GSPO, use `sequence`, default is None which corresponds to\n # the original GRPO paper.\n importance_sampling_level: Literal['sequence', 'token'] | None\n\n # Whether to sync the reference model.\n sync_ref_model: bool | None = False\n # Mixup alpha for the reference model.\n ref_model_mixup_alpha: float | None = 0.9\n # Sync steps for the reference model.\n ref_model_sync_steps: int | None = 64\n # Whether to scale rewards by their standard deviation.\n scale_rewards: bool = True\n\n # Sampling temperature for the GRPO policy.\n temperature: float | None\n # Top-p sampling probability for the generation policy.\n top_p: float | None\n # Top-k sampling for the generation policy.\n top_k: int | None\n # Minimum probability for the generation policy.\n min_p: float | None\n # Penalty for tokens that appear in prompt and generated text.\n repetition_penalty: float | None\n # Number of iterations per batch (μ) for GRPO.\n num_iterations: int | None\n # Epsilon value for clipping in the GRPO algorithm.\n epsilon: float | None\n # Upper-bound epsilon value for clipping in the GRPO algorithm.\n epsilon_high: float | None\n # Whether to use Liger loss for GRPO.\n use_liger_loss: bool | None\n # Loss formulation to use. Supported values: grpo, bnpo, dr_grpo.\n loss_type: str | None\n # Whether to exclude truncated completions from loss calculation.\n mask_truncated_completions: bool = False\n\nvllm: VllmConfig | None\n # For VllmConfig:\n # Device to use for VLLM\n device: str | None = auto\n # Tensor parallel size for VLLM\n tensor_parallel_size: int | None\n # Data parallel size for VLLM\n data_parallel_size: int | None\n # GPU memory utilization for VLLM\n gpu_memory_utilization: float | None = 0.9\n # Data type for VLLM\n dtype: str | None = auto\n # Maximum length of the model context for VLLM\n max_model_len: int | None\n # Enable prefix caching for VLLM\n enable_prefix_caching: bool | None\n # Host for the vLLM server to start on\n host: str | None = 0.0.0.0\n # Port of the vLLM server to start on\n port: int | None = 8000\n\n # Enable reasoning for VLLM\n enable_reasoning: bool | None\n # Reasoning parser for VLLM\n reasoning_parser: str | None\n\nqat: QATConfig | None\n # For QATConfig:\n # Fake quantization layout to use for activation quantization. Valid options are\n # \"int4\" and \"int8\"\n activation_dtype: TorchIntDType | None\n # Fake quantization layout to use for weight quantization. Valid options are \"int4\"\n # and \"int8\"\n weight_dtype: TorchIntDType = TorchIntDType.int8\n # Quantize embedding\n quantize_embedding: bool | None = False\n # The number of elements in each group for per-group fake quantization\n group_size: int | None = 32\n # The number of steps to apply fake quantization after\n fake_quant_after_n_steps: int | None\n\nquantization: PTQConfig | None\n # For PTQConfig:\n # Fake quantization layout to use for weight quantization. Valid options are uintX for\n # X in [1, 2, 3, 4, 5, 6, 7], or int4, or int8\n weight_dtype: TorchIntDType = TorchIntDType.int8\n # Fake quantization layout to use for activation quantization. Valid options are\n # \"int4\" and \"int8\"\n activation_dtype: TorchIntDType | None\n # Whether to quantize the embedding layer.\n quantize_embedding: bool | None\n # The number of elements in each group for per-group fake quantization\n group_size: int | None = 32\n\n# Reward modelling: `True` or `False`\nreward_model: bool | None\n# Process reward modelling: `True` or `False`\nprocess_reward_model: bool | None\nnum_labels: int | None\n\n# Whether to perform weighting in DPO trainer\ndpo_use_weighting: bool | None\ndpo_use_logits_to_keep: bool | None\ndpo_label_smoothing: float | None\ndpo_norm_loss: bool | None\ndpo_padding_free: bool | None\ndpo_generate_during_eval: bool | None\n\n# A list of one or more datasets to finetune the model with\ndatasets: Annotated[list[SFTDataset | DPODataset | KTODataset | StepwiseSupervisedDataset], MinLen(1)] | None\n # For SFTDataset:\n # HuggingFace dataset repo | s3:// | gs:// | path to local file or directory\n path: str | None\n # name of dataset split to load from\n split: str | None\n # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]\n type: str | UserDefinedPrompterType | None\n # For UserDefinedPrompterType:\n # Custom user instruction prompt\n system_prompt: str | None\n # Use {system} as key to be replaced\n system_format: str | None\n field_system: str | None\n field_instruction: str | None\n field_input: str | None\n field_output: str | None\n\n # Customizable to be single line or multi-line. Use {instruction}/{input} as key to\n # be replaced. 'format' can include {input}\n format: str | None\n # 'no_input_format' cannot include {input}\n no_input_format: str | None\n input_transform: str | None\n # split dataset into N pieces (use with shards_idx)\n shards: int | None\n # the index of sharded dataset to use\n shards_idx: int | None\n # process dataset in N sequential chunks for memory efficiency (exclusive with\n # `shards`)\n preprocess_shards: int | None\n conversation: str | None\n\n # The name of the chat template to use for training, following values are supported:\n # tokenizer_default: Uses the chat template that is available in the\n # tokenizer_config.json. If the chat template is not available in the tokenizer, it\n # will raise an error. This is the default.\n # alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates\n # are available in the axolotl codebase at src/axolotl/utils/chat_templates.py.\n # tokenizer_default_fallback_*: where * is the name of the chat template to fallback\n # to if the tokenizer does not have a chat template else default to tokenizer. E.g.\n # tokenizer_default_fallback_chatml. jinja: Uses a custom jinja template for the chat\n # template. The custom jinja template should be provided in the chat_template_jinja\n # field.\n chat_template: ChatTemplate | str | None\n # Custom jinja chat template or path to jinja file. Used only if `chat_template:\n # jinja` or empty.\n chat_template_jinja: str | None\n # path to source data files\n data_files: str | list[str] | None\n input_format: str | None\n # name of dataset configuration to load\n name: str | None\n # defines the datatype when path is a file\n ds_type: str | None\n # For `completion` datasets only, uses the provided field instead of `text` column\n field: str | None\n field_human: str | None\n field_model: str | None\n # Key containing the messages (default: \"messages\")\n field_messages: str | None\n # Key containing the tools (default: \"tools\"). Must be a list[dict] and follow [JSON\n # schema](https://json-schema.org/learn/getting-started-step-by-step).\n field_tools: str | None\n # Key containing the reasoning trace (default: \"reasoning_content\").\n field_thinking: str | None\n # The key the chat template expects that indicates the reasoning trace.\n template_thinking_key: str | None\n\n message_field_role: str | None\n\n message_field_content: str | None\n # Mapping of properties from the input dataset to the chat template. (default:\n # message_property_mappings={'role':'role', 'content':'content'}) If a property exists\n # in the template but not in this mapping, the system will attempt to load it directly\n # from the message using the property name as the key. Example: In the mapping below,\n # 'from' is loaded from input dataset and used as 'role', while 'value' is loaded and\n # used as 'content' in the chat template.\n message_property_mappings: dict[str, str] | None\n # The key in the message turn that indicates via boolean whether tokens of a turn\n # should be considered for training. Useful to selectively train on certain turns\n # besides the `roles_to_train`.\n message_field_training: str | None\n # The key in the message turn that contains the training details. Useful to\n # selectively train on certain tokens in a turn. The value of the key is a List[Dict]\n # containing `begin_offset` (start character index in content), `end_offset` (end\n # character index in content), and `train` (boolean whether to train).\n message_field_training_detail: str | None\n # (for Qwen3 template only) Whether to split the assistant content based on a\n # reasoning trace inside delimited tags\n split_thinking: bool | None\n logprobs_field: str | None\n temperature: float | None\n # Roles to train on. The tokens from these roles will be considered for the loss.\n roles_to_train: list[str] | None\n # Which EOS tokens to train on in the conversation. Possible values are: all: train on\n # all EOS tokens, turn (default): train on the EOS token at the end of each trainable\n # turn, last: train on the last EOS token in the conversation\n train_on_eos: Literal['all', 'turn', 'last'] | None\n # Roles mapping in the messages. The format is {target_role: [source_roles]}. All\n # source roles will be mapped to the target role. The default is: user: [\"human\",\n # \"user\"], assistant: [\"gpt\", \"assistant\"], system: [\"system\"], tool: [\"tool\"]\n roles: dict[str, list[str]] | None\n # Whether to drop the system turn from the dataset. Only works with chat_template.\n # This does not drop the default system message from chat_template if it exists. If\n # you wish to, we recommend using a custom jinja template with the default system\n # message removed or adding a system turn with empty content.\n drop_system_message: bool | None\n # Trust remote code for untrusted source\n trust_remote_code: bool | None = False\n # The specific revision of the dataset to use when loading from the Hugging Face Hub.\n # This can be a commit hash, tag, or branch name. If not specified, the latest version\n # will be used. This parameter is ignored for local datasets.\n revision: str | None\n\n # For DPODataset:\n path: str | None\n split: str | None\n type: UserDefinedDPOType | str | None\n # For UserDefinedDPOType:\n field_system: str | None\n field_prompt: str | None\n field_chosen: str | None\n field_rejected: str | None\n prompt_format: str | None\n chosen_format: str | None\n rejected_format: str | None\n data_files: list[str] | None\n revision: str | None\n field_messages: str | None\n\n # For KTODataset:\n path: str | None\n split: str | None\n type: UserDefinedKTOType | str | None\n # For UserDefinedKTOType:\n field_system: str | None\n field_prompt: str | None\n field_completion: str | None\n field_label: bool | None\n prompt_format: str | None\n completion_format: str | None\n data_files: list[str] | None\n trust_remote_code: bool | None = False\n revision: str | None\n\n # For StepwiseSupervisedDataset:\n path: str | None\n split: str | None\n data_files: list[str] | None\n revision: str | None\n step_separator: str | None\n max_completion_length: int | None\n train_on_last_step_only: bool | None\n\n# A list of one or more datasets to eval the model with. You can use either\n# test_datasets, or val_set_size, but not both.\ntest_datasets: Annotated[list[SFTDataset | DPODataset | KTODataset | StepwiseSupervisedDataset], MinLen(1)] | None\n # For SFTDataset:\n # HuggingFace dataset repo | s3:// | gs:// | path to local file or directory\n path: str | None\n # name of dataset split to load from\n split: str | None\n # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]\n type: str | UserDefinedPrompterType | None\n # For UserDefinedPrompterType:\n # Custom user instruction prompt\n system_prompt: str | None\n # Use {system} as key to be replaced\n system_format: str | None\n field_system: str | None\n field_instruction: str | None\n field_input: str | None\n field_output: str | None\n\n # Customizable to be single line or multi-line. Use {instruction}/{input} as key to\n # be replaced. 'format' can include {input}\n format: str | None\n # 'no_input_format' cannot include {input}\n no_input_format: str | None\n input_transform: str | None\n # split dataset into N pieces (use with shards_idx)\n shards: int | None\n # the index of sharded dataset to use\n shards_idx: int | None\n # process dataset in N sequential chunks for memory efficiency (exclusive with\n # `shards`)\n preprocess_shards: int | None\n conversation: str | None\n\n # The name of the chat template to use for training, following values are supported:\n # tokenizer_default: Uses the chat template that is available in the\n # tokenizer_config.json. If the chat template is not available in the tokenizer, it\n # will raise an error. This is the default.\n # alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates\n # are available in the axolotl codebase at src/axolotl/utils/chat_templates.py.\n # tokenizer_default_fallback_*: where * is the name of the chat template to fallback\n # to if the tokenizer does not have a chat template else default to tokenizer. E.g.\n # tokenizer_default_fallback_chatml. jinja: Uses a custom jinja template for the chat\n # template. The custom jinja template should be provided in the chat_template_jinja\n # field.\n chat_template: ChatTemplate | str | None\n # Custom jinja chat template or path to jinja file. Used only if `chat_template:\n # jinja` or empty.\n chat_template_jinja: str | None\n # path to source data files\n data_files: str | list[str] | None\n input_format: str | None\n # name of dataset configuration to load\n name: str | None\n # defines the datatype when path is a file\n ds_type: str | None\n # For `completion` datasets only, uses the provided field instead of `text` column\n field: str | None\n field_human: str | None\n field_model: str | None\n # Key containing the messages (default: \"messages\")\n field_messages: str | None\n # Key containing the tools (default: \"tools\"). Must be a list[dict] and follow [JSON\n # schema](https://json-schema.org/learn/getting-started-step-by-step).\n field_tools: str | None\n # Key containing the reasoning trace (default: \"reasoning_content\").\n field_thinking: str | None\n # The key the chat template expects that indicates the reasoning trace.\n template_thinking_key: str | None\n\n message_field_role: str | None\n\n message_field_content: str | None\n # Mapping of properties from the input dataset to the chat template. (default:\n # message_property_mappings={'role':'role', 'content':'content'}) If a property exists\n # in the template but not in this mapping, the system will attempt to load it directly\n # from the message using the property name as the key. Example: In the mapping below,\n # 'from' is loaded from input dataset and used as 'role', while 'value' is loaded and\n # used as 'content' in the chat template.\n message_property_mappings: dict[str, str] | None\n # The key in the message turn that indicates via boolean whether tokens of a turn\n # should be considered for training. Useful to selectively train on certain turns\n # besides the `roles_to_train`.\n message_field_training: str | None\n # The key in the message turn that contains the training details. Useful to\n # selectively train on certain tokens in a turn. The value of the key is a List[Dict]\n # containing `begin_offset` (start character index in content), `end_offset` (end\n # character index in content), and `train` (boolean whether to train).\n message_field_training_detail: str | None\n # (for Qwen3 template only) Whether to split the assistant content based on a\n # reasoning trace inside delimited tags\n split_thinking: bool | None\n logprobs_field: str | None\n temperature: float | None\n # Roles to train on. The tokens from these roles will be considered for the loss.\n roles_to_train: list[str] | None\n # Which EOS tokens to train on in the conversation. Possible values are: all: train on\n # all EOS tokens, turn (default): train on the EOS token at the end of each trainable\n # turn, last: train on the last EOS token in the conversation\n train_on_eos: Literal['all', 'turn', 'last'] | None\n # Roles mapping in the messages. The format is {target_role: [source_roles]}. All\n # source roles will be mapped to the target role. The default is: user: [\"human\",\n # \"user\"], assistant: [\"gpt\", \"assistant\"], system: [\"system\"], tool: [\"tool\"]\n roles: dict[str, list[str]] | None\n # Whether to drop the system turn from the dataset. Only works with chat_template.\n # This does not drop the default system message from chat_template if it exists. If\n # you wish to, we recommend using a custom jinja template with the default system\n # message removed or adding a system turn with empty content.\n drop_system_message: bool | None\n # Trust remote code for untrusted source\n trust_remote_code: bool | None = False\n # The specific revision of the dataset to use when loading from the Hugging Face Hub.\n # This can be a commit hash, tag, or branch name. If not specified, the latest version\n # will be used. This parameter is ignored for local datasets.\n revision: str | None\n\n # For DPODataset:\n path: str | None\n split: str | None\n type: UserDefinedDPOType | str | None\n # For UserDefinedDPOType:\n field_system: str | None\n field_prompt: str | None\n field_chosen: str | None\n field_rejected: str | None\n prompt_format: str | None\n chosen_format: str | None\n rejected_format: str | None\n data_files: list[str] | None\n revision: str | None\n field_messages: str | None\n\n # For KTODataset:\n path: str | None\n split: str | None\n type: UserDefinedKTOType | str | None\n # For UserDefinedKTOType:\n field_system: str | None\n field_prompt: str | None\n field_completion: str | None\n field_label: bool | None\n prompt_format: str | None\n completion_format: str | None\n data_files: list[str] | None\n trust_remote_code: bool | None = False\n revision: str | None\n\n # For StepwiseSupervisedDataset:\n path: str | None\n split: str | None\n data_files: list[str] | None\n revision: str | None\n step_separator: str | None\n max_completion_length: int | None\n train_on_last_step_only: bool | None\n\n# If false, the datasets will not be shuffled and will keep their original order in\n# `datasets`. The same applies to the `test_datasets` option and the\n# `pretraining_dataset` option. Default is true.\nshuffle_merged_datasets: bool | None = True\n# If true, each dataset in `datasets` will be shuffled before merging. This allows\n# curriculum learning strategies to be applied at the dataset level. Default is false.\nshuffle_before_merging_datasets: bool | None = False\n# Axolotl attempts to save the dataset as an arrow after packing the data together so\n# subsequent training attempts load faster, relative path\ndataset_prepared_path: str | None\n# Num shards for whole dataset\ndataset_shard_num: int | None\n# Index of shard to use for whole dataset\ndataset_shard_idx: int | None\nskip_prepare_dataset: bool | None = False\n# Number of shards to save the prepared dataset\nnum_dataset_shards_to_save: int | None\n\n# Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize\npretraining_dataset: Annotated[list[PretrainingDataset | SFTDataset], MinLen(1)] | None\n # For PretrainingDataset:\n name: str | None\n path: str | None\n split: str | None = train\n text_column: str | None = text\n type: str | None = pretrain\n trust_remote_code: bool | None = False\n data_files: str | None\n skip: int | None\n\n # For SFTDataset:\n # HuggingFace dataset repo | s3:// | gs:// | path to local file or directory\n path: str | None\n # name of dataset split to load from\n split: str | None\n # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]\n type: str | UserDefinedPrompterType | None\n # For UserDefinedPrompterType:\n # Custom user instruction prompt\n system_prompt: str | None\n # Use {system} as key to be replaced\n system_format: str | None\n field_system: str | None\n field_instruction: str | None\n field_input: str | None\n field_output: str | None\n\n # Customizable to be single line or multi-line. Use {instruction}/{input} as key to\n # be replaced. 'format' can include {input}\n format: str | None\n # 'no_input_format' cannot include {input}\n no_input_format: str | None\n input_transform: str | None\n # split dataset into N pieces (use with shards_idx)\n shards: int | None\n # the index of sharded dataset to use\n shards_idx: int | None\n # process dataset in N sequential chunks for memory efficiency (exclusive with\n # `shards`)\n preprocess_shards: int | None\n conversation: str | None\n\n # The name of the chat template to use for training, following values are supported:\n # tokenizer_default: Uses the chat template that is available in the\n # tokenizer_config.json. If the chat template is not available in the tokenizer, it\n # will raise an error. This is the default.\n # alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates\n # are available in the axolotl codebase at src/axolotl/utils/chat_templates.py.\n # tokenizer_default_fallback_*: where * is the name of the chat template to fallback\n # to if the tokenizer does not have a chat template else default to tokenizer. E.g.\n # tokenizer_default_fallback_chatml. jinja: Uses a custom jinja template for the chat\n # template. The custom jinja template should be provided in the chat_template_jinja\n # field.\n chat_template: ChatTemplate | str | None\n # Custom jinja chat template or path to jinja file. Used only if `chat_template:\n # jinja` or empty.\n chat_template_jinja: str | None\n # path to source data files\n data_files: str | list[str] | None\n input_format: str | None\n # name of dataset configuration to load\n name: str | None\n # defines the datatype when path is a file\n ds_type: str | None\n # For `completion` datasets only, uses the provided field instead of `text` column\n field: str | None\n field_human: str | None\n field_model: str | None\n # Key containing the messages (default: \"messages\")\n field_messages: str | None\n # Key containing the tools (default: \"tools\"). Must be a list[dict] and follow [JSON\n # schema](https://json-schema.org/learn/getting-started-step-by-step).\n field_tools: str | None\n # Key containing the reasoning trace (default: \"reasoning_content\").\n field_thinking: str | None\n # The key the chat template expects that indicates the reasoning trace.\n template_thinking_key: str | None\n\n message_field_role: str | None\n\n message_field_content: str | None\n # Mapping of properties from the input dataset to the chat template. (default:\n # message_property_mappings={'role':'role', 'content':'content'}) If a property exists\n # in the template but not in this mapping, the system will attempt to load it directly\n # from the message using the property name as the key. Example: In the mapping below,\n # 'from' is loaded from input dataset and used as 'role', while 'value' is loaded and\n # used as 'content' in the chat template.\n message_property_mappings: dict[str, str] | None\n # The key in the message turn that indicates via boolean whether tokens of a turn\n # should be considered for training. Useful to selectively train on certain turns\n # besides the `roles_to_train`.\n message_field_training: str | None\n # The key in the message turn that contains the training details. Useful to\n # selectively train on certain tokens in a turn. The value of the key is a List[Dict]\n # containing `begin_offset` (start character index in content), `end_offset` (end\n # character index in content), and `train` (boolean whether to train).\n message_field_training_detail: str | None\n # (for Qwen3 template only) Whether to split the assistant content based on a\n # reasoning trace inside delimited tags\n split_thinking: bool | None\n logprobs_field: str | None\n temperature: float | None\n # Roles to train on. The tokens from these roles will be considered for the loss.\n roles_to_train: list[str] | None\n # Which EOS tokens to train on in the conversation. Possible values are: all: train on\n # all EOS tokens, turn (default): train on the EOS token at the end of each trainable\n # turn, last: train on the last EOS token in the conversation\n train_on_eos: Literal['all', 'turn', 'last'] | None\n # Roles mapping in the messages. The format is {target_role: [source_roles]}. All\n # source roles will be mapped to the target role. The default is: user: [\"human\",\n # \"user\"], assistant: [\"gpt\", \"assistant\"], system: [\"system\"], tool: [\"tool\"]\n roles: dict[str, list[str]] | None\n # Whether to drop the system turn from the dataset. Only works with chat_template.\n # This does not drop the default system message from chat_template if it exists. If\n # you wish to, we recommend using a custom jinja template with the default system\n # message removed or adding a system turn with empty content.\n drop_system_message: bool | None\n # Trust remote code for untrusted source\n trust_remote_code: bool | None = False\n # The specific revision of the dataset to use when loading from the Hugging Face Hub.\n # This can be a commit hash, tag, or branch name. If not specified, the latest version\n # will be used. This parameter is ignored for local datasets.\n revision: str | None\n\n# The maximum number of processes to use while preprocessing your input dataset. This\n# defaults to `os.cpu_count()` if not set. For Runpod VMs, it will default to number of\n# vCPUs via RUNPOD_CPU_COUNT.\ndataset_processes: int | None\n# Deduplicates datasets and test_datasets with identical entries\ndataset_exact_deduplication: bool | None\n# Keep dataset in memory while preprocessing. Only needed if cached dataset is taking\n# too much storage\ndataset_keep_in_memory: bool | None\ndataloader_pin_memory: bool | None\ndataloader_num_workers: int | None\ndataloader_prefetch_factor: int | None\ndataloader_drop_last: bool | None\n\naccelerator_config: dict[str, Any] | None\n\nremove_unused_columns: bool | None\n\n# Push prepared dataset to hub - repo_org/repo_name\npush_dataset_to_hub: str | None\n# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private\n# datasets. Required to be true when used in combination with `push_dataset_to_hub`\nhf_use_auth_token: bool | None\n\ndevice: Any | None\n# Passed through to transformers when loading the model when launched without\n# accelerate. Use `sequential` when training w/ model parallelism to limit memory\ndevice_map: Any | None\nworld_size: int | None\n# Don't mess with this, it's here for accelerate and torchrun\nlocal_rank: int | None\nddp: bool | None\n\n# Seed for reproducibility\nseed: int | None\n# Advanced DDP Arguments - timeout\nddp_timeout: int | None\n# Advanced DDP Arguments - bucket cap in MB\nddp_bucket_cap_mb: int | None\n# Advanced DDP Arguments - broadcast buffers\nddp_broadcast_buffers: bool | None\nddp_find_unused_parameters: bool | None\n\n# Approximate number of predictions sent to wandb depending on batch size. Enabled above\n# 0. Default is 0\neval_table_size: int | None\n# Total number of tokens generated for predictions sent to wandb. Default is 128\neval_max_new_tokens: int | None\n# Whether to run causal language model evaluation for metrics in\n# `eval_causal_lm_metrics`\ndo_causal_lm_eval: bool | None\n# HF evaluate metrics used during evaluation. Default is ['sacrebleu', 'comet', 'ter',\n# 'chrf', 'perplexity']\neval_causal_lm_metrics: list[str] | None\ndo_bench_eval: bool | None\nbench_dataset: str | None\nbench_split: str | None\nmetric_for_best_model: str | None\ngreater_is_better: bool | None\n\n# High loss value, indicating the learning has broken down (a good estimate is ~2 times\n# the loss at the start of training)\nloss_watchdog_threshold: float | None\n# Number of high-loss steps in a row before the trainer aborts (default: 3)\nloss_watchdog_patience: int | None\n\n# Run garbage collection every `gc_steps` steps. -1 will run on epoch end and before\n# evaluations. Default is 0 (disabled).\ngc_steps: int | None\n\n# Use CUDA bf16. bool or 'full' for `bf16_full_eval`, or 'auto' for automatic detection.\n# require >=ampere\nbf16: Literal['auto'] | bool | None = auto\n# Use CUDA fp16\nfp16: bool | None\n# Enable FP8 mixed precision training using TorchAO. Best used in combination with\n# torch.compile.\nfp8: bool | None\n# Enable FSDP float8 all-gather optimization for FP8 training. Can improve training\n# speed by 10-15% when FSDP is enabled.\nfp8_enable_fsdp_float8_all_gather: bool | None\n# No AMP (automatic mixed precision) - require >=ampere\nbfloat16: bool | None\n# No AMP (automatic mixed precision)\nfloat16: bool | None\n# Use CUDA tf32 - require >=ampere\ntf32: bool | None\nfloat32: bool | None\n\n# Whether to use gradient checkpointing. Available options are: true, false, 'offload',\n# 'offload_disk'.\n# https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing\ngradient_checkpointing: Literal['offload', 'offload_disk'] | bool | None = False\n# Additional kwargs to pass to the trainer for gradient checkpointing\ngradient_checkpointing_kwargs: dict[str, Any] | None\n# Whether to offload activations. Available options are: true, false, 'legacy', 'disk'.\nactivation_offloading: Literal['legacy', 'disk'] | bool | None = False\n\nunfrozen_parameters: list[str] | None\n\n# The maximum length of an input to train with, this should typically be less than 2048\n# as most models have a token/context limit of 2048\nsequence_len: int = 512\n# The maximum length of an input for evaluation. If not specified, defaults to\n# sequence_len\neval_sequence_len: int | None\nmin_sample_len: int | None\n# maximum prompt length for RL training\nmax_prompt_len: int = 512\n# Use efficient multi-packing with block diagonal attention and per sequence\n# position_ids. Recommend set to 'true'\nsample_packing: bool | None\n# The number of samples packed at a time. Increasing the following values helps with\n# packing, but usually only slightly (<%1.)\nsample_packing_group_size: int | None = 100000\n# The number of samples which can be packed into one sequence. Increase if using a large\n# sequence_len with many short samples.\nsample_packing_bin_size: int | None = 200\n# Whether to pack samples sequentially\nsample_packing_sequentially: bool | None\n# The multiprocessing start method to use for packing. Should be 'fork', 'spawn' or\n# 'forkserver'\nsample_packing_mp_start_method: str | None\n# Set to 'false' if getting errors during eval with sample_packing on\neval_sample_packing: bool | None\n# Pad inputs so each step uses constant sized buffers. This will reduce memory\n# fragmentation and may prevent OOMs, by re-using memory more efficiently. Defaults to\n# True if `sample_packing` enabled\npad_to_sequence_len: bool | None\n# Whether to use sequential sampling for curriculum learning\ncurriculum_sampling: bool | None\nmultipack_real_batches: bool | None\n# whether to concatenate samples during pretraining\npretraining_sample_concatenation: bool | None\n\n# Use batch flattening for speedups when not using sample_packing\nbatch_flattening: Literal['auto'] | bool | None\n\nuse_pose: bool | None\npose_split_on_token_ids: list[int] | None\npose_max_context_len: int | None\npose_num_chunks: int | None\n\npretrain_multipack_buffer_size: int | None = 10000\n# whether to prevent cross attention for packed sequences during pretraining\npretrain_multipack_attn: bool | None = True\n\n# Whether to use xformers attention patch https://github.com/facebookresearch/xformers\nxformers_attention: bool | None\n# Whether to use scaled-dot-product attention https://pytorch.org/docs/stable/generated/\n# torch.nn.functional.scaled_dot_product_attention.html\nsdp_attention: bool | None\n# Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf\ns2_attention: bool | None\nflex_attention: bool | None\nflex_attn_compile_kwargs: dict[str, Any] | None\n# Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention\nflash_attention: bool | None\n# Whether to use flash-attention cross entropy implementation - advanced use only\nflash_attn_cross_entropy: bool | None\n# Whether to use flash-attention rms norm implementation - advanced use only\nflash_attn_rms_norm: bool | None\n# Whether to fuse part of the MLP into a single operation\nflash_attn_fuse_mlp: bool | None\n# Whether to use bettertransformers\nflash_optimum: bool | None\n\neager_attention: bool | None\n\n# Specify a custom attention implementation, used mostly for kernels.\nattn_implementation: str | None\n\nunsloth_cross_entropy_loss: bool | None\nunsloth_lora_mlp: bool | None\nunsloth_lora_qkv: bool | None\nunsloth_lora_o: bool | None\nunsloth_rms_norm: bool | None\nunsloth_rope: bool | None\n\n# Apply custom LoRA autograd functions and activation function Triton kernels for speed\n# and memory savings. See: https://docs.axolotl.ai/docs/lora_optims.html\nlora_mlp_kernel: bool | None\n# Apply custom LoRA autograd functions and activation function Triton kernels for speed\n# and memory savings. See: https://docs.axolotl.ai/docs/lora_optims.html\nlora_qkv_kernel: bool | None\n# Apply custom LoRA autograd functions and activation function Triton kernels for speed\n# and memory savings. See: https://docs.axolotl.ai/docs/lora_optims.html\nlora_o_kernel: bool | None\n\n# Whether to use chunked cross entropy loss for memory efficiency\nchunked_cross_entropy: bool | None\n# Number of chunks to use for chunked cross entropy loss\nchunked_cross_entropy_num_chunks: int | None\n\n# Whether to use ALST tiled mlp for memory efficient long context\ntiled_mlp: bool | None\n\n# Number of shards to use for ALST tiled mlp. If unset, it will be set based on\n# seqlen/hidden_size\ntiled_mlp_num_shards: int | None\n\n# Whether to use original mlp for ALST tiled mlp. Otherwise uses a generic MLP based on\n# llama.\ntiled_mlp_use_original_mlp: bool | None = True\n\nllama4_linearized_experts: bool | None\n\n# Deepspeed config path. e.g., deepspeed_configs/zero3.json\ndeepspeed: str | dict[str, Any] | None\n# Whether to use deepcompile for faster training with deepspeed\ndeepcompile: bool | None\n# FSDP configuration\nfsdp: list[str] | None\n\n# FSDP configuration options\nfsdp_config: dict[str, Any] | None\n# FSDP version\nfsdp_version: int | None\nfsdp_final_state_dict_type: Literal['FULL_STATE_DICT', 'LOCAL_STATE_DICT', 'SHARDED_STATE_DICT'] | None\n\n# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for\n# no eval.\nval_set_size: float | None = 0.0\n\n# Number of devices to shard across. If not set, will use all available devices.\ndp_shard_size: int | None\n# Number of devices to replicate across.\ndp_replicate_size: int | None\n# Deprecated: use `context_parallel_size` instead\nsequence_parallel_degree: int | None\n# Set to a divisor of the number of GPUs available to split sequences into chunks of\n# equal size. Use in long context training to prevent OOM when sequences cannot fit into\n# a single GPU's VRAM. E.g., if 4 GPUs are available, set this value to 2 to split each\n# sequence into two equal-sized subsequences, or set to 4 to split into four equal-sized\n# subsequences. See https://docs.axolotl.ai/docs/sequence_parallelism.html for more\n# details.\ncontext_parallel_size: int | None\n# Optional; strides across the key dimension. Larger values use more memory but should\n# make training faster. Must evenly divide the number of KV heads in your model.\nheads_k_stride: int | None\n# One of 'varlen_llama3', 'batch_ring', 'batch_zigzag', 'batch_stripe'. Defaults to\n# 'varlen_llama3' in the sample packing case, and 'batch_ring' in the non-sample packing\n# case.\nring_attn_func: RingAttnFunc | None\n# Number of tensor parallel processes in TP group. Only supported with DeepSpeed AutoTP.\ntensor_parallel_size: int | None\n\n# Add or change special tokens. If you add tokens here, you don't need to add them to\n# the `tokens` list.\nspecial_tokens: SpecialTokensConfig | None\n # For SpecialTokensConfig:\n bos_token: str | None\n eos_token: str | None\n pad_token: str | None\n unk_token: str | None\n additional_special_tokens: list[str] | None\n\n# Add extra tokens to the tokenizer\ntokens: list[str] | None\n# Mapping token_id to new_token_string to override reserved added_tokens in the\n# tokenizer. Only works for tokens that are not part of the base vocab (aka are\n# added_tokens). Can be checked if they exist in tokenizer.json added_tokens.\nadded_tokens_overrides: dict[int, str] | None\n\n# Whether to use torch.compile and which backend to use. setting to `auto` will enable\n# torch compile when torch>=2.6.0\ntorch_compile: Literal['auto'] | bool | None\n# Backend to use for torch.compile\ntorch_compile_backend: str | None\ntorch_compile_mode: Literal['default', 'reduce-overhead', 'max-autotune'] | None\n\n# Maximum number of iterations to train for. It precedes num_epochs which means that if\n# both are set, num_epochs will not be guaranteed. e.g., when 1 epoch is 1000 steps =>\n# `num_epochs: 2` and `max_steps: 100` will train for 100 steps\nmax_steps: int | None\n# Number of warmup steps. Cannot use with warmup_ratio\nwarmup_steps: int | None\n# Warmup ratio. Cannot use with warmup_steps\nwarmup_ratio: float | None\n# Leave empty to eval at each epoch, integer for every N steps. float for fraction of\n# total steps\neval_steps: int | float | None\n# Number of times per epoch to run evals, mutually exclusive with eval_steps\nevals_per_epoch: int | None\n# Set to `no` to skip evaluation, `epoch` at end of each epoch, leave empty to infer\n# from `eval_steps`\neval_strategy: str | None\n\n# Leave empty to save at each epoch, integer for every N steps. float for fraction of\n# total steps\nsave_steps: int | float | None\n# Number of times per epoch to save a checkpoint, mutually exclusive with save_steps\nsaves_per_epoch: int | None\n# Set to `no` to skip checkpoint saves, `epoch` at end of each epoch, `best` when better\n# result is achieved, leave empty to infer from `save_steps`\nsave_strategy: str | None\n# Checkpoints saved at a time\nsave_total_limit: int | None\n# Whether to checkpoint a model after the first step of training. Defaults to False.\nsave_first_step: bool | None\n\n# Logging frequency\nlogging_steps: int | None\n# Stop training after this many evaluation losses have increased in a row. https://huggi\n# ngface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppin\n# gCallback\nearly_stopping_patience: int | None\nload_best_model_at_end: bool | None = False\n# Save only the model weights, skipping the optimizer. Using this means you can't resume\n# from checkpoints.\nsave_only_model: bool | None = False\n# Use tensorboard for logging\nuse_tensorboard: bool | None\n# Enable the pytorch profiler to capture the first N steps of training to the\n# output_dir. see https://pytorch.org/blog/understanding-gpu-memory-1/ for more\n# information. Snapshots can be visualized @ https://pytorch.org/memory_viz\nprofiler_steps: int | None\n# Which step to start the profiler at. Useful for only capturing a few steps mid-run.\nprofiler_steps_start: int | None = 0\n# bool of whether to include tokens trainer per second in the training metrics. This\n# iterates over the entire dataset once, so it takes some time.\ninclude_tokens_per_second: bool | None\n\n# NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to\n# add noise to embeddings. Currently only supported on Llama and Mistral\nneftune_noise_alpha: float | None\n\n# Parameter controlling the relative ratio loss weight in the ORPO loss. Passed to\n# `beta` in `ORPOConfig` due to trl mapping.\norpo_alpha: float | None\n# Weighting of NLL term in loss from RPO paper\nrpo_alpha: float | None\n# Target reward margin for the SimPO loss\nsimpo_gamma: float | None\n# Weight of the BC regularizer\ncpo_alpha: float | None\n\n# Factor for desirable loss term in KTO loss\nkto_desirable_weight: float | None\n# Factor for undesirable loss term in KTO loss\nkto_undesirable_weight: float | None\n# The beta parameter for the RL training\nrl_beta: float | None\n\n# Defines the max memory usage per gpu on the system. Passed through to transformers\n# when loading the model.\nmax_memory: dict[int | Literal['cpu', 'disk'], int | str] | None\n# Limit the memory for all available GPUs to this amount (if an integer, expressed in\n# gigabytes); default: unset\ngpu_memory_limit: int | str | None\n# Whether to use low_cpu_mem_usage\nlow_cpu_mem_usage: bool | None\n\n# The name of the chat template to use for training, following values are supported:\n# tokenizer_default: Uses the chat template that is available in the\n# tokenizer_config.json. If the chat template is not available in the tokenizer, it will\n# raise an error. This is the default value.\n# alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates\n# are available in the axolotl codebase at src/axolotl/utils/chat_templates.py.\n# tokenizer_default_fallback_*: where * is the name of the chat template to fallback to.\n# E.g. tokenizer_default_fallback_chatml. This is useful when the chat template is not\n# available in the tokenizer. jinja: Uses a custom jinja template for the chat template.\n# The custom jinja template should be provided in the chat_template_jinja field. The\n# selected chat template will be saved to the tokenizer_config.json for easier\n# inferencing\nchat_template: ChatTemplate | Annotated[str, StringConstraints(pattern='^tokenizer_default_fallback_')] | None\n# Custom jinja template or path to jinja file for chat template. This will be only used\n# if chat_template is set to `jinja` or `null` (in which case chat_template is\n# automatically set to `jinja`). Default is null.\nchat_template_jinja: str | None\n# Additional kwargs to pass to the chat template. This is useful for customizing the\n# chat template. For example, you can pass `thinking=False` to add a generation prompt\n# to the chat template.\nchat_template_kwargs: dict[str, Any] | None\n# Custom EOT (End-of-Turn) tokens to mask/unmask during training. These tokens mark the\n# boundaries between conversation turns. For example: ['/INST', '</s>',\n# '[/SYSTEM_PROMPT]']. If not specified, defaults to just the model's eos_token. This is\n# useful for templates that use multiple delimiter tokens.\neot_tokens: list[str] | None\n# Changes the default system message. Currently only supports chatml.\ndefault_system_message: str | None\n\nfix_untrained_tokens: int | list[int] | None\n\nis_preprocess: bool | None\npreprocess_iterable: bool | None\n\n# Total number of tokens - internal use\ntotal_num_tokens: int | None\ntotal_supervised_tokens: int | None\n# You can set these packing optimizations AFTER starting a training at least once. The\n# trainer will provide recommended values for these values.\nsample_packing_eff_est: float | None\naxolotl_config_path: str | None\n\n# Internal use only - Used to identify which the model is based on\nis_falcon_derived_model: bool | None\n# Internal use only - Used to identify which the model is based on\nis_llama_derived_model: bool | None\n# Internal use only - Used to identify which the model is based on. Please note that if\n# you set this to true, `padding_side` will be set to 'left' by default\nis_mistral_derived_model: bool | None\n# Internal use only - Used to identify which the model is based on\nis_qwen_derived_model: bool | None\n\n# Add plugins to extend the pipeline. See `src/axolotl/integrations` for the available\n# plugins or doc below for more details.\n# https://docs.axolotl.ai/docs/custom_integrations.html\nplugins: list[str] | None\n\n# This is the huggingface model that contains *.pt, *.safetensors, or *.bin files. This\n# can also be a relative path to a model on disk\nbase_model: str (required)\n# If the base_model repo on hf hub doesn't include configuration .json files, You can\n# set that here, or leave this empty to default to base_model\nbase_model_config: str | None\ncls_model_config: str | None\n# Optional tokenizer configuration path in case you want to use a different tokenizer\n# than the one defined in the base model\ntokenizer_config: str | None\n# use_fast option for tokenizer loading from_pretrained, default to True\ntokenizer_use_fast: bool | None\n# Whether to use the legacy tokenizer setting, defaults to True\ntokenizer_legacy: bool | None\n# Whether to use mistral-common tokenizer. If set to True, it will use the mistral-\n# common tokenizer.\ntokenizer_use_mistral_common: bool | None\n# Corresponding tokenizer for the model AutoTokenizer is a good choice\ntokenizer_type: str | None\n# transformers processor class\nprocessor_type: str | None\n# Trust remote code for untrusted source\ntrust_remote_code: bool | None\n\n# Don't move the model to the device before sharding. This is an experimental feature\n# that may be included in the future as the default.\nexperimental_skip_move_to_device: bool | None\n\n# Use custom kernels, e.g. MegaBlocks.\nuse_kernels: bool | None\n\n# Model loading quantization config\nmodel_quantization_config: Literal['Mxfp4Config'] | None\n# kwargs for model quantization config\nmodel_quantization_config_kwargs: dict[str, Any] | None\n\n# Where to save the full-finetuned model to\noutput_dir: str = ./model-out\n# push checkpoints to hub\nhub_model_id: str | None\n# how to push checkpoints to hub\nhub_strategy: str | None\n# Save model as safetensors (require safetensors package). Default True\nsave_safetensors: bool | None = True\n\n# This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer\nload_in_8bit: bool | None = False\n# Use bitsandbytes 4 bit\nload_in_4bit: bool | None = False\n\n# If you want to use 'lora' or 'qlora' or leave blank to train all parameters in\n# original model\nadapter: str | None\n# If you already have a lora model trained that you want to load, put that here. This\n# means after training, if you want to test the model, you should set this to the value\n# of `output_dir`. Note that if you merge an adapter to the base model, a new\n# subdirectory `merged` will be created under the `output_dir`.\nlora_model_dir: str | None\nlora_r: int | None\nlora_alpha: int | None\nlora_fan_in_fan_out: bool | None\nlora_target_modules: str | list[str] | None\nlora_target_parameters: str | list[str] | None\n# If true, will target all linear modules\nlora_target_linear: bool | None\n# If you added new tokens to the tokenizer, you may need to save some LoRA modules\n# because they need to know the new tokens. For LLaMA and Mistral, you need to save\n# `embed_tokens` and `lm_head`. It may vary for other models. `embed_tokens` converts\n# tokens to embeddings, and `lm_head` converts embeddings to token probabilities.\nlora_modules_to_save: list[str] | None\nlora_dropout: float | None = 0.0\n# The layer indices to transform, otherwise, apply to all layers\npeft_layers_to_transform: list[int] | None\npeft_layers_pattern: list[str] | None\n\npeft: PeftConfig | None\n # For PeftConfig:\n # Configuration options for loftq initialization for LoRA\n loftq_config: LoftQConfig | None\n # For LoftQConfig:\n # typically 4 bits\n loftq_bits: int = 4\n\n# Whether to use DoRA.\npeft_use_dora: bool | None\n# Whether to use RSLoRA.\npeft_use_rslora: bool | None\n# List of layer indices to replicate.\npeft_layer_replication: list[tuple[int, int]] | None\n# How to initialize LoRA weights. Default to True which is MS original implementation.\npeft_init_lora_weights: bool | str | None\n\n# load qlora model in sharded format for FSDP using answer.ai technique.\nqlora_sharded_model_loading: bool | None = False\n# Do the LoRA/PEFT loading on CPU -- this is required if the base model is so large it\n# takes up most or all of the available GPU VRAM, e.g. during a model and LoRA merge\nlora_on_cpu: bool | None\n# Whether you are training a 4-bit GPTQ quantized model\ngptq: bool | None\n# optional overrides to the bnb 4bit quantization configuration\nbnb_config_kwargs: dict[str, Any] | None\n\n# loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4.\nloraplus_lr_ratio: float | None\n# loraplus learning rate for lora embedding layers. Default value is 1e-6.\nloraplus_lr_embedding: float | None = 1e-06\n\nmerge_lora: bool | None\n\n# Whether to use ReLoRA. Use with jagged_restart_*steps options.\nrelora: bool | None\n# threshold for optimizer magnitude when pruning\nrelora_prune_ratio: float | None\n# True to perform lora weight merges on cpu during restarts, for modest gpu memory\n# savings\nrelora_cpu_offload: bool | None\n\n# how often to reset for jagged restarts\njagged_restart_steps: int | None\n# how many warmup steps to take after reset for jagged restarts\njagged_restart_warmup_steps: int | None\n# how many anneal steps to take before reset for jagged restarts\njagged_restart_anneal_steps: int | None\n\n# If greater than 1, backpropagation will be skipped and the gradients will be\n# accumulated for the given number of steps.\ngradient_accumulation_steps: int | None = 1\n# The number of samples to include in each batch. This is the number of samples sent to\n# each GPU. Batch size per gpu = micro_batch_size * gradient_accumulation_steps\nmicro_batch_size: int | None = 1\n# Total batch size, we do not recommended setting this manually\nbatch_size: int | None\n# per gpu micro batch size for evals, defaults to value of micro_batch_size\neval_batch_size: int | None\n\n# whether to find batch size that fits in memory. Passed to underlying transformers\n# Trainer\nauto_find_batch_size: bool | None\n\n# Whether to mask out or include the human's prompt from the training labels\ntrain_on_inputs: bool | None = False\n# Group similarly sized data to minimize padding. May be slower to start, as it must\n# download and sort the entire dataset. Note that training loss may have an oscillating\n# pattern with this enabled.\ngroup_by_length: bool | None\n\nlearning_rate: str | float (required)\nembedding_lr: float | None\nembedding_lr_scale: float | None\n# Specify weight decay\nweight_decay: float | None = 0.0\n# Specify optimizer\noptimizer: OptimizerNames | CustomSupportedOptimizers | None = OptimizerNames.ADAMW_TORCH_FUSED\n# Dictionary of arguments to pass to the optimizer\noptim_args: str | dict[str, Any] | None\n# The target modules to optimize, i.e. the module names that you would like to train,\n# right now this is used only for GaLore algorithm\noptim_target_modules: list[str] | Literal['all_linear'] | None\n# Path to torch distx for optim 'adamw_anyprecision'\ntorchdistx_path: str | None\nlr_scheduler: SchedulerType | Literal['one_cycle'] | Literal['rex'] | None = SchedulerType.COSINE\n# Specify a scheduler and kwargs to use with the optimizer\nlr_scheduler_kwargs: dict[str, Any] | None\nlr_quadratic_warmup: bool | None\n# decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of\n# peak lr\ncosine_min_lr_ratio: float | None\n# freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means\n# start cosine_min_lr at 80% of training step\ncosine_constant_lr_ratio: float | None\n# Learning rate div factor\nlr_div_factor: float | None\n\nlr_groups: list[LrGroup] | None\n # For LrGroup:\n name: str (required)\n modules: list[str] (required)\n lr: float (required)\n\n# adamw hyperparams\nadam_epsilon: float | None\n# only used for CAME Optimizer\nadam_epsilon2: float | None\n# adamw hyperparams\nadam_beta1: float | None\n# adamw hyperparams\nadam_beta2: float | None\n# only used for CAME Optimizer\nadam_beta3: float | None\n\n# Dion Optimizer learning rate\ndion_lr: float | None\n# Dion Optimizer momentum\ndion_momentum: float | None\n# Dion Optimizer: r/d fraction for low-rank approximation. Used to compute the low-rank\n# dimension.\ndion_rank_fraction: float | None = 1.0\n# Dion Optimizer: Round up the low-rank dimension to a multiple of this number. This may\n# be useful to ensure even sharding.\ndion_rank_multiple_of: int | None = 1\n\n# Gradient clipping max norm\nmax_grad_norm: float | None\nnum_epochs: float = 1.0\n\nuse_wandb: bool | None\n# Set the name of your wandb run\nwandb_name: str | None\n# Set the ID of your wandb run\nwandb_run_id: str | None\n# \"offline\" to save run metadata locally and not sync to the server, \"disabled\" to turn\n# off wandb\nwandb_mode: str | None\n# Your wandb project name\nwandb_project: str | None\n# A wandb Team name if using a Team\nwandb_entity: str | None\nwandb_watch: str | None\n# \"checkpoint\" to log model to wandb Artifacts every `save_steps` or \"end\" to log only\n# at the end of training\nwandb_log_model: str | None\n\nuse_mlflow: bool | None\n# URI to mlflow\nmlflow_tracking_uri: str | None\n# Your experiment name\nmlflow_experiment_name: str | None\n# Your run name\nmlflow_run_name: str | None\n# set to true to copy each saved checkpoint on each save to mlflow artifact registry\nhf_mlflow_log_artifacts: bool | None\n\n# Enable or disable Comet integration.\nuse_comet: bool | None\n# API key for Comet. Recommended to set via `comet login`.\ncomet_api_key: str | None\n# Workspace name in Comet. Defaults to the user's default workspace.\ncomet_workspace: str | None\n# Project name in Comet. Defaults to Uncategorized.\ncomet_project_name: str | None\n# Identifier for the experiment. Used to append data to an existing experiment or\n# control the key of new experiments. Default to a random key.\ncomet_experiment_key: str | None\n# Create a new experiment (\"create\") or log to an existing one (\"get\"). Default\n# (\"get_or_create\") auto-selects based on configuration.\ncomet_mode: str | None\n# Set to True to log data to Comet server, or False for offline storage. Default is\n# True.\ncomet_online: bool | None\n# Dictionary for additional configuration settings, see the doc for more details.\ncomet_experiment_config: dict[str, Any] | None\n\n# the number of activate layers in LISA\nlisa_n_layers: int | None\n# how often to switch layers in LISA\nlisa_step_interval: int | None\n# path under the model to access the layers\nlisa_layers_attribute: str | None = model.layers\n\ngradio_title: str | None\ngradio_share: bool | None\ngradio_server_name: str | None\ngradio_server_port: int | None\ngradio_max_new_tokens: int | None\ngradio_temperature: float | None\n\nuse_ray: bool = False\nray_run_name: str | None\nray_num_workers: int = 1\nresources_per_worker: dict\n\n# The size of the image to resize to. It can be an integer (resized into padded-square\n# image) or a tuple (width, height).If not provided, we will attempt to load from\n# preprocessor.size, otherwise, images won't be resized.\nimage_size: int | tuple[int, int] | None\n# The resampling algorithm to use for image resizing. Default is bilinear. Please refer\n# to PIL.Image.Resampling for more details.\nimage_resize_algorithm: Literal['bilinear', 'bicubic', 'lanczos'] | Resampling | None\n\n# optional overrides to the base model configuration\noverrides_of_model_config: dict[str, Any] | None\n# optional overrides the base model loading from_pretrained\noverrides_of_model_kwargs: dict[str, Any] | None\n# If you want to specify the type of model to load, AutoModelForCausalLM is a good\n# choice too\ntype_of_model: str | None\n# You can specify to choose a specific model revision from huggingface hub\nrevision_of_model: str | None\n\nmax_packed_sequence_len: int | None\nrope_scaling: Any | None\nnoisy_embedding_alpha: float | None\ndpo_beta: float | None\nevaluation_strategy: str | None", "crumbs": [ "Getting Started", "Config Reference" @@ -2467,7 +2467,7 @@ "href": "docs/dataset-formats/conversation.html", "title": "Conversation", "section": "", - "text": "Chat Template strategy uses a jinja2 template that converts a list of messages into a prompt. Support using tokenizer’s template, a supported template, or custom jinja2.\n\n\ndata.jsonl\n\n{\"messages\": [{\"role\": \"...\", \"content\": \"...\"}, {\"role\": \"...\", \"content\": \"...\"}, ...]}\n\nSee configs for full configs and supported templates.\n\n\nMost configs can be adapted as follows:\n# old\nchat_template: chatml\ndatasets:\n - path: ...\n type: sharegpt\n conversation: chatml\n\n# new (if using tokenizer's chat_template)\ndatasets:\n - path: ...\n type: chat_template\n\n field_messages: conversations\n message_property_mappings:\n role: from\n content: value\n\n# new (if setting a new chat_template like chatml, gemma, etc)\nchat_template: chatml\ndatasets:\n - path: ...\n type: chat_template\n\n field_messages: conversations\n message_property_mappings:\n role: from\n content: value\nWe recommend checking the below examples for other usecases.\n\n\n\n\n\n(Legacy) Using the default chat template in the tokenizer_config.json on OpenAI messages format, training on only last message.\ndatasets:\n - path: ...\n type: chat_template\n roles_to_train:\n train_on_eos:\n\n\n\n\n\n\nTip\n\n\n\nIf you receive an error like “chat_template choice is tokenizer_default but tokenizer’s chat_template is null.”, it means the tokenizer does not have a default chat_template. Follow the examples below instead to set a custom chat_template.\n\n\n\n\n\nUsing the gemma chat template to override the tokenizer_config.json’s chat template on OpenAI messages format, training on all assistant messages.\nchat_template: gemma # this overwrites the tokenizer's chat_template\ndatasets:\n - path: ...\n type: chat_template\n roles_to_train: [\"assistant\"] # default value\n\n\n\n\n\n\nNote\n\n\n\nIf you want to use built-in chat_template, use chat_template: tokenizer_default (this is set by default).\n\n\n\n\n\nUsing the tokenizer_config.json’s chat template or chatml as fallback if the former’s chat template does not exist, on OpenAI messages format, training on all assistant messages.\nchat_template: tokenizer_default_fallback_chatml # this overwrites the tokenizer's chat_template\ndatasets:\n - path: ...\n type: chat_template\n\n\n\nUsing a custom jinja template on OpenAI messages format, training on all assistant messages.\n# chat_template: jinja # `jinja` will be implied if the `chat_template_jinja` is set and this field is empty\nchat_template_jinja: \"{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'system') %}{{'<|system|>' + '\\n' + message['content'] + '<|end|>' + '\\n'}}{% elif (message['role'] == 'user') %}{{'<|user|>' + '\\n' + message['content'] + '<|end|>' + '\\n' + '<|assistant|>' + '\\n'}}{% elif message['role'] == 'assistant' %}{{message['content'] + '<|end|>' + '\\n'}}{% endif %}{% endfor %}\"\n\ndatasets:\n - path: ...\n type: chat_template\n\n\n\n\n\n\nImportant\n\n\n\nPlease make sure that your tokenizer.eos_token is same as EOS (End-of-Sequence) token in template. Otherwise, set eos_token under special_tokens:.\n\n\n\n\n\n\nIf you are using a template that has a different EOT (End-of-Turn) token from EOS token or multiple EOT tokens (like Mistral V7 Tekken), set the eot_tokens: config. The handling of EOT tokens follows train_on_eos: which defaults to turn.\n\neot_tokens:\n - \"[/INST]\"\n # - \"[/SYSTEM_PROMPT]\"\n\ndatasets:\n - path: ...\n type: chat_template\n\n # optional\n train_on_eot: turn # defaults read from train_on_eos (which defaults to turn)\n\n\n\n\n\n\nTip\n\n\n\nSee config documentation for detailed explanations of “turn”, “last”, and “all” options for training on tokens.\n\n\n\n\n\n\n\n\nNote\n\n\n\nUsing eot_tokens requires each token that exists in chat_template to be a single token in the tokenizer. Otherwise, the tokenizer will split the token and cause unexpected behavior.\nYou can add those tokens as new tokens under tokens: or (recommended) override unused added_tokens via added_tokens_overrides:. See config for more details.\n\n\n\nContinuing from the previous example, if you want to train on all EOT token trainable turns but only last EOS token, set train_on_eos: last.\n\neot_tokens:\n - \"[/INST]\"\n # ...\n\ndatasets:\n - path: ...\n type: chat_template\n\n train_on_eos: last\n train_on_eot: turn\n\n\n\n\n\n\nTip\n\n\n\nIf EOS token only appears at the end of a prompt, train_on_eos: last is equivalent to train_on_eos: turn. Therefore, generally, you can leave them to their defaults and omit them.\n\n\n\n\n\nInstead of passing tools via the system prompt, an alternative method would be to have the tools in a separate column and loaded via chat_template to let the template dynamically build it.\n{\n \"tools\": [\n {\n \"type\": \"...\",\n \"function\": {\n \"name\": \"...\",\n \"description\": \"...\",\n \"parameters\": {\n \"type\": \"...\",\n \"properties\": {\n // ...\n },\n \"required\": [\"...\"],\n },\n },\n },\n ],\n \"messages\": [\n // ...\n {\n \"role\": \"assistant\", // call the function via assistant\n \"tool_calls\": [\n {\n \"id\": \"...\", // required only for mistral\n \"type\": \"function\",\n \"function\": {\n \"name\": \"...\",\n \"arguments\": {\n \"...\": \"...\",\n }\n }\n }\n ]\n },\n {\n \"role\": \"tool\",\n \"tool_call_id\": \"...\", // required only for mistral\n \"name\": \"...\",\n \"content\": \"...\"\n },\n ],\n}\n\n\n\n\n\n\nNote\n\n\n\nTools need to follow JSON schema.\n\n\nchat_template: llama4\ndatasets:\n - path: ...\n type: chat_template\n # field_tools: tools # default is `tools`\n\n\n\n\n\n\nTip\n\n\n\nLook into the chat_template you are using to see if it supports tools and what the expected role is for the tool answer. In the example above, the tool answer is expected to be in the tool or ipython role for llama4 template.\n\n\n\n\n\n(Advanced) Using fine-grained control over tokens and turns to train in a conversation\nFor a data sample that looks like:\n\n\ndata.jsonl\n\n{\n \"conversations\": [\n {\"from\": \"system\", \"value\": \"You are an AI assistant.\", \"train\": false},\n {\"from\": \"human\", \"value\": \"Hello\", \"train\": false},\n {\"from\": \"assistant\", \"value\": \"Hello\", \"train\": true},\n {\"from\": \"human\", \"value\": \"How are you?\", \"train\": true},\n {\n \"from\": \"assistant\",\n \"value\": \"I'm doing very well, thank you!\",\n \"train_detail\": [\n {\"begin_offset\": 0, \"end_offset\": 8, \"train\": false},\n {\"begin_offset\": 9, \"end_offset\": 18, \"train\": true},\n {\"begin_offset\": 19, \"end_offset\": 30, \"train\": false},\n ],\n },\n {\n \"from\": \"human\",\n \"value\": \"I'm doing very well, thank you!\",\n \"train\": true,\n },\n {\"from\": \"assistant\", \"value\": \"Hi there!\", \"train\": true}\n ]\n}\n\nThe configuration would look like:\ndatasets:\n - path: ...\n type: chat_template\n chat_template: tokenizer_default\n field_messages: conversations\n message_property_mappings:\n role: from\n content: value\n roles_to_train: []\n train_on_eos: turn\n message_field_training: train\n message_field_training_detail: train_detail\n\n\n\n\n\n\nTip\n\n\n\nIt is not necessary to set both message_field_training and message_field_training_detail at once.\n\n\n\n\n\n(For Qwen3 template only) Enable reasoning split, where the reasoning is split from the content and passed as a separate field into the template.\ndatasets:\n - path: ...\n type: chat_template\n chat_template: qwen3\n split_thinking: true\nFor example, a content can look like:\n{\n \"content\": \"<think>Some thinking outputs</think>Output after thinking.\"\n}\nAfter split, it will look like:\n{\n \"reasoning_content\": \"Some thinking outputs\",\n \"content\": \"Output after thinking...\"\n}", + "text": "Chat Template strategy uses a jinja2 template that converts a list of messages into a prompt. Support using tokenizer’s template, a supported template, or custom jinja2.\n\n\ndata.jsonl\n\n{\"messages\": [{\"role\": \"...\", \"content\": \"...\"}, {\"role\": \"...\", \"content\": \"...\"}, ...]}\n\nSee configs for full configs and supported templates.\n\n\nMost configs can be adapted as follows:\n# old\nchat_template: chatml\ndatasets:\n - path: ...\n type: sharegpt\n conversation: chatml\n\n# new (if using tokenizer's chat_template)\ndatasets:\n - path: ...\n type: chat_template\n\n field_messages: conversations\n message_property_mappings:\n role: from\n content: value\n\n# new (if setting a new chat_template like chatml, gemma, etc)\nchat_template: chatml\ndatasets:\n - path: ...\n type: chat_template\n\n field_messages: conversations\n message_property_mappings:\n role: from\n content: value\nWe recommend checking the below examples for other usecases.\n\n\n\n\n\n(Legacy) Using the default chat template in the tokenizer_config.json on OpenAI messages format, training on only last message.\ndatasets:\n - path: ...\n type: chat_template\n roles_to_train:\n train_on_eos:\n\n\n\n\n\n\nTip\n\n\n\nIf you receive an error like “chat_template choice is tokenizer_default but tokenizer’s chat_template is null.”, it means the tokenizer does not have a default chat_template. Follow the examples below instead to set a custom chat_template.\n\n\n\n\n\nUsing the gemma chat template to override the tokenizer_config.json’s chat template on OpenAI messages format, training on all assistant messages.\nchat_template: gemma # this overwrites the tokenizer's chat_template\ndatasets:\n - path: ...\n type: chat_template\n roles_to_train: [\"assistant\"] # default value\n\n\n\n\n\n\nNote\n\n\n\nIf you want to use built-in chat_template, use chat_template: tokenizer_default (this is set by default).\n\n\n\n\n\nUsing the tokenizer_config.json’s chat template or chatml as fallback if the former’s chat template does not exist, on OpenAI messages format, training on all assistant messages.\nchat_template: tokenizer_default_fallback_chatml # this overwrites the tokenizer's chat_template\ndatasets:\n - path: ...\n type: chat_template\n\n\n\nUsing a custom jinja template on OpenAI messages format, training on all assistant messages.\n# chat_template: jinja # `jinja` will be implied if the `chat_template_jinja` is set and this field is empty\nchat_template_jinja: \"{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'system') %}{{'<|system|>' + '\\n' + message['content'] + '<|end|>' + '\\n'}}{% elif (message['role'] == 'user') %}{{'<|user|>' + '\\n' + message['content'] + '<|end|>' + '\\n' + '<|assistant|>' + '\\n'}}{% elif message['role'] == 'assistant' %}{{message['content'] + '<|end|>' + '\\n'}}{% endif %}{% endfor %}\"\n\ndatasets:\n - path: ...\n type: chat_template\n\n\n\n\n\n\nImportant\n\n\n\nPlease make sure that your tokenizer.eos_token is same as EOS (End-of-Sequence) token in template. Otherwise, set eos_token under special_tokens:.\n\n\n\n\n\n\nIf you are using a template that has a different EOT (End-of-Turn) token from EOS token or multiple EOT tokens (like Mistral V7 Tekken), set the eot_tokens: config. The handling of EOT tokens follows train_on_eos: which defaults to turn.\n\neot_tokens:\n - \"[/INST]\"\n # - \"[/SYSTEM_PROMPT]\"\n\ndatasets:\n - path: ...\n type: chat_template\n\n # optional\n train_on_eot: turn # defaults read from train_on_eos (which defaults to turn)\n\n\n\n\n\n\nTip\n\n\n\nSee config documentation for detailed explanations of “turn”, “last”, and “all” options for training on tokens.\n\n\n\n\n\n\n\n\nNote\n\n\n\nUsing eot_tokens requires each token that exists in chat_template to be a single token in the tokenizer. Otherwise, the tokenizer will split the token and cause unexpected behavior.\nYou can add those tokens as new tokens under tokens: or (recommended) override unused added_tokens via added_tokens_overrides:. See config for more details.\n\n\n\nContinuing from the previous example, if you want to train on all EOT token trainable turns but only last EOS token, set train_on_eos: last.\n\neot_tokens:\n - \"[/INST]\"\n # ...\n\ndatasets:\n - path: ...\n type: chat_template\n\n train_on_eos: last\n train_on_eot: turn\n\n\n\n\n\n\nTip\n\n\n\nIf EOS token only appears at the end of a prompt, train_on_eos: last is equivalent to train_on_eos: turn. Therefore, generally, you can leave them to their defaults and omit them.\n\n\n\n\n\nInstead of passing tools via the system prompt, an alternative method would be to have the tools in a separate column and loaded via chat_template to let the template dynamically build it.\n{\n \"tools\": [\n {\n \"type\": \"...\",\n \"function\": {\n \"name\": \"...\",\n \"description\": \"...\",\n \"parameters\": {\n \"type\": \"...\",\n \"properties\": {\n // ...\n },\n \"required\": [\"...\"],\n },\n },\n },\n ],\n \"messages\": [\n // ...\n {\n \"role\": \"assistant\", // call the function via assistant\n \"tool_calls\": [\n {\n \"id\": \"...\", // required only for mistral\n \"type\": \"function\",\n \"function\": {\n \"name\": \"...\",\n \"arguments\": {\n \"...\": \"...\",\n }\n }\n }\n ]\n },\n {\n \"role\": \"tool\",\n \"tool_call_id\": \"...\", // required only for mistral\n \"name\": \"...\",\n \"content\": \"...\"\n },\n ],\n}\n\n\n\n\n\n\nNote\n\n\n\nTools need to follow JSON schema.\n\n\nExample config for Llama4:\nchat_template: llama4\ndatasets:\n - path: Nanobit/text-tools-2k-test\n type: chat_template\n # field_tools: tools # default is `tools`\n\n\n\n\n\n\nTip\n\n\n\nLook into the chat_template you are using to see if it supports tools and what the expected role is for the tool answer. In the example above, the tool answer is expected to be in the tool or ipython role for llama4 template.\n\n\n\n\n\n(Advanced) Using fine-grained control over tokens and turns to train in a conversation\nFor a data sample that looks like:\n\n\ndata.jsonl\n\n{\n \"conversations\": [\n {\"from\": \"system\", \"value\": \"You are an AI assistant.\", \"train\": false},\n {\"from\": \"human\", \"value\": \"Hello\", \"train\": false},\n {\"from\": \"assistant\", \"value\": \"Hello\", \"train\": true},\n {\"from\": \"human\", \"value\": \"How are you?\", \"train\": true},\n {\n \"from\": \"assistant\",\n \"value\": \"I'm doing very well, thank you!\",\n \"train_detail\": [\n {\"begin_offset\": 0, \"end_offset\": 8, \"train\": false},\n {\"begin_offset\": 9, \"end_offset\": 18, \"train\": true},\n {\"begin_offset\": 19, \"end_offset\": 30, \"train\": false},\n ],\n },\n {\n \"from\": \"human\",\n \"value\": \"I'm doing very well, thank you!\",\n \"train\": true,\n },\n {\"from\": \"assistant\", \"value\": \"Hi there!\", \"train\": true}\n ]\n}\n\nThe configuration would look like:\ndatasets:\n - path: ...\n type: chat_template\n chat_template: tokenizer_default\n field_messages: conversations\n message_property_mappings:\n role: from\n content: value\n roles_to_train: []\n train_on_eos: turn\n message_field_training: train\n message_field_training_detail: train_detail\n\n\n\n\n\n\nTip\n\n\n\nIt is not necessary to set both message_field_training and message_field_training_detail at once.\n\n\n\n\n\n(For Qwen3 template only) Enable reasoning split, where the reasoning is split from the content and passed as a separate field into the template.\ndatasets:\n - path: ...\n type: chat_template\n chat_template: qwen3\n split_thinking: true\nFor example, a content can look like:\n{\n \"content\": \"<think>Some thinking outputs</think>Output after thinking.\"\n}\nAfter split, it will look like:\n{\n \"reasoning_content\": \"Some thinking outputs\",\n \"content\": \"Output after thinking...\"\n}", "crumbs": [ "Dataset Formats", "Conversation" @@ -2478,7 +2478,7 @@ "href": "docs/dataset-formats/conversation.html#chat_template", "title": "Conversation", "section": "", - "text": "Chat Template strategy uses a jinja2 template that converts a list of messages into a prompt. Support using tokenizer’s template, a supported template, or custom jinja2.\n\n\ndata.jsonl\n\n{\"messages\": [{\"role\": \"...\", \"content\": \"...\"}, {\"role\": \"...\", \"content\": \"...\"}, ...]}\n\nSee configs for full configs and supported templates.\n\n\nMost configs can be adapted as follows:\n# old\nchat_template: chatml\ndatasets:\n - path: ...\n type: sharegpt\n conversation: chatml\n\n# new (if using tokenizer's chat_template)\ndatasets:\n - path: ...\n type: chat_template\n\n field_messages: conversations\n message_property_mappings:\n role: from\n content: value\n\n# new (if setting a new chat_template like chatml, gemma, etc)\nchat_template: chatml\ndatasets:\n - path: ...\n type: chat_template\n\n field_messages: conversations\n message_property_mappings:\n role: from\n content: value\nWe recommend checking the below examples for other usecases.\n\n\n\n\n\n(Legacy) Using the default chat template in the tokenizer_config.json on OpenAI messages format, training on only last message.\ndatasets:\n - path: ...\n type: chat_template\n roles_to_train:\n train_on_eos:\n\n\n\n\n\n\nTip\n\n\n\nIf you receive an error like “chat_template choice is tokenizer_default but tokenizer’s chat_template is null.”, it means the tokenizer does not have a default chat_template. Follow the examples below instead to set a custom chat_template.\n\n\n\n\n\nUsing the gemma chat template to override the tokenizer_config.json’s chat template on OpenAI messages format, training on all assistant messages.\nchat_template: gemma # this overwrites the tokenizer's chat_template\ndatasets:\n - path: ...\n type: chat_template\n roles_to_train: [\"assistant\"] # default value\n\n\n\n\n\n\nNote\n\n\n\nIf you want to use built-in chat_template, use chat_template: tokenizer_default (this is set by default).\n\n\n\n\n\nUsing the tokenizer_config.json’s chat template or chatml as fallback if the former’s chat template does not exist, on OpenAI messages format, training on all assistant messages.\nchat_template: tokenizer_default_fallback_chatml # this overwrites the tokenizer's chat_template\ndatasets:\n - path: ...\n type: chat_template\n\n\n\nUsing a custom jinja template on OpenAI messages format, training on all assistant messages.\n# chat_template: jinja # `jinja` will be implied if the `chat_template_jinja` is set and this field is empty\nchat_template_jinja: \"{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'system') %}{{'<|system|>' + '\\n' + message['content'] + '<|end|>' + '\\n'}}{% elif (message['role'] == 'user') %}{{'<|user|>' + '\\n' + message['content'] + '<|end|>' + '\\n' + '<|assistant|>' + '\\n'}}{% elif message['role'] == 'assistant' %}{{message['content'] + '<|end|>' + '\\n'}}{% endif %}{% endfor %}\"\n\ndatasets:\n - path: ...\n type: chat_template\n\n\n\n\n\n\nImportant\n\n\n\nPlease make sure that your tokenizer.eos_token is same as EOS (End-of-Sequence) token in template. Otherwise, set eos_token under special_tokens:.\n\n\n\n\n\n\nIf you are using a template that has a different EOT (End-of-Turn) token from EOS token or multiple EOT tokens (like Mistral V7 Tekken), set the eot_tokens: config. The handling of EOT tokens follows train_on_eos: which defaults to turn.\n\neot_tokens:\n - \"[/INST]\"\n # - \"[/SYSTEM_PROMPT]\"\n\ndatasets:\n - path: ...\n type: chat_template\n\n # optional\n train_on_eot: turn # defaults read from train_on_eos (which defaults to turn)\n\n\n\n\n\n\nTip\n\n\n\nSee config documentation for detailed explanations of “turn”, “last”, and “all” options for training on tokens.\n\n\n\n\n\n\n\n\nNote\n\n\n\nUsing eot_tokens requires each token that exists in chat_template to be a single token in the tokenizer. Otherwise, the tokenizer will split the token and cause unexpected behavior.\nYou can add those tokens as new tokens under tokens: or (recommended) override unused added_tokens via added_tokens_overrides:. See config for more details.\n\n\n\nContinuing from the previous example, if you want to train on all EOT token trainable turns but only last EOS token, set train_on_eos: last.\n\neot_tokens:\n - \"[/INST]\"\n # ...\n\ndatasets:\n - path: ...\n type: chat_template\n\n train_on_eos: last\n train_on_eot: turn\n\n\n\n\n\n\nTip\n\n\n\nIf EOS token only appears at the end of a prompt, train_on_eos: last is equivalent to train_on_eos: turn. Therefore, generally, you can leave them to their defaults and omit them.\n\n\n\n\n\nInstead of passing tools via the system prompt, an alternative method would be to have the tools in a separate column and loaded via chat_template to let the template dynamically build it.\n{\n \"tools\": [\n {\n \"type\": \"...\",\n \"function\": {\n \"name\": \"...\",\n \"description\": \"...\",\n \"parameters\": {\n \"type\": \"...\",\n \"properties\": {\n // ...\n },\n \"required\": [\"...\"],\n },\n },\n },\n ],\n \"messages\": [\n // ...\n {\n \"role\": \"assistant\", // call the function via assistant\n \"tool_calls\": [\n {\n \"id\": \"...\", // required only for mistral\n \"type\": \"function\",\n \"function\": {\n \"name\": \"...\",\n \"arguments\": {\n \"...\": \"...\",\n }\n }\n }\n ]\n },\n {\n \"role\": \"tool\",\n \"tool_call_id\": \"...\", // required only for mistral\n \"name\": \"...\",\n \"content\": \"...\"\n },\n ],\n}\n\n\n\n\n\n\nNote\n\n\n\nTools need to follow JSON schema.\n\n\nchat_template: llama4\ndatasets:\n - path: ...\n type: chat_template\n # field_tools: tools # default is `tools`\n\n\n\n\n\n\nTip\n\n\n\nLook into the chat_template you are using to see if it supports tools and what the expected role is for the tool answer. In the example above, the tool answer is expected to be in the tool or ipython role for llama4 template.\n\n\n\n\n\n(Advanced) Using fine-grained control over tokens and turns to train in a conversation\nFor a data sample that looks like:\n\n\ndata.jsonl\n\n{\n \"conversations\": [\n {\"from\": \"system\", \"value\": \"You are an AI assistant.\", \"train\": false},\n {\"from\": \"human\", \"value\": \"Hello\", \"train\": false},\n {\"from\": \"assistant\", \"value\": \"Hello\", \"train\": true},\n {\"from\": \"human\", \"value\": \"How are you?\", \"train\": true},\n {\n \"from\": \"assistant\",\n \"value\": \"I'm doing very well, thank you!\",\n \"train_detail\": [\n {\"begin_offset\": 0, \"end_offset\": 8, \"train\": false},\n {\"begin_offset\": 9, \"end_offset\": 18, \"train\": true},\n {\"begin_offset\": 19, \"end_offset\": 30, \"train\": false},\n ],\n },\n {\n \"from\": \"human\",\n \"value\": \"I'm doing very well, thank you!\",\n \"train\": true,\n },\n {\"from\": \"assistant\", \"value\": \"Hi there!\", \"train\": true}\n ]\n}\n\nThe configuration would look like:\ndatasets:\n - path: ...\n type: chat_template\n chat_template: tokenizer_default\n field_messages: conversations\n message_property_mappings:\n role: from\n content: value\n roles_to_train: []\n train_on_eos: turn\n message_field_training: train\n message_field_training_detail: train_detail\n\n\n\n\n\n\nTip\n\n\n\nIt is not necessary to set both message_field_training and message_field_training_detail at once.\n\n\n\n\n\n(For Qwen3 template only) Enable reasoning split, where the reasoning is split from the content and passed as a separate field into the template.\ndatasets:\n - path: ...\n type: chat_template\n chat_template: qwen3\n split_thinking: true\nFor example, a content can look like:\n{\n \"content\": \"<think>Some thinking outputs</think>Output after thinking.\"\n}\nAfter split, it will look like:\n{\n \"reasoning_content\": \"Some thinking outputs\",\n \"content\": \"Output after thinking...\"\n}", + "text": "Chat Template strategy uses a jinja2 template that converts a list of messages into a prompt. Support using tokenizer’s template, a supported template, or custom jinja2.\n\n\ndata.jsonl\n\n{\"messages\": [{\"role\": \"...\", \"content\": \"...\"}, {\"role\": \"...\", \"content\": \"...\"}, ...]}\n\nSee configs for full configs and supported templates.\n\n\nMost configs can be adapted as follows:\n# old\nchat_template: chatml\ndatasets:\n - path: ...\n type: sharegpt\n conversation: chatml\n\n# new (if using tokenizer's chat_template)\ndatasets:\n - path: ...\n type: chat_template\n\n field_messages: conversations\n message_property_mappings:\n role: from\n content: value\n\n# new (if setting a new chat_template like chatml, gemma, etc)\nchat_template: chatml\ndatasets:\n - path: ...\n type: chat_template\n\n field_messages: conversations\n message_property_mappings:\n role: from\n content: value\nWe recommend checking the below examples for other usecases.\n\n\n\n\n\n(Legacy) Using the default chat template in the tokenizer_config.json on OpenAI messages format, training on only last message.\ndatasets:\n - path: ...\n type: chat_template\n roles_to_train:\n train_on_eos:\n\n\n\n\n\n\nTip\n\n\n\nIf you receive an error like “chat_template choice is tokenizer_default but tokenizer’s chat_template is null.”, it means the tokenizer does not have a default chat_template. Follow the examples below instead to set a custom chat_template.\n\n\n\n\n\nUsing the gemma chat template to override the tokenizer_config.json’s chat template on OpenAI messages format, training on all assistant messages.\nchat_template: gemma # this overwrites the tokenizer's chat_template\ndatasets:\n - path: ...\n type: chat_template\n roles_to_train: [\"assistant\"] # default value\n\n\n\n\n\n\nNote\n\n\n\nIf you want to use built-in chat_template, use chat_template: tokenizer_default (this is set by default).\n\n\n\n\n\nUsing the tokenizer_config.json’s chat template or chatml as fallback if the former’s chat template does not exist, on OpenAI messages format, training on all assistant messages.\nchat_template: tokenizer_default_fallback_chatml # this overwrites the tokenizer's chat_template\ndatasets:\n - path: ...\n type: chat_template\n\n\n\nUsing a custom jinja template on OpenAI messages format, training on all assistant messages.\n# chat_template: jinja # `jinja` will be implied if the `chat_template_jinja` is set and this field is empty\nchat_template_jinja: \"{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'system') %}{{'<|system|>' + '\\n' + message['content'] + '<|end|>' + '\\n'}}{% elif (message['role'] == 'user') %}{{'<|user|>' + '\\n' + message['content'] + '<|end|>' + '\\n' + '<|assistant|>' + '\\n'}}{% elif message['role'] == 'assistant' %}{{message['content'] + '<|end|>' + '\\n'}}{% endif %}{% endfor %}\"\n\ndatasets:\n - path: ...\n type: chat_template\n\n\n\n\n\n\nImportant\n\n\n\nPlease make sure that your tokenizer.eos_token is same as EOS (End-of-Sequence) token in template. Otherwise, set eos_token under special_tokens:.\n\n\n\n\n\n\nIf you are using a template that has a different EOT (End-of-Turn) token from EOS token or multiple EOT tokens (like Mistral V7 Tekken), set the eot_tokens: config. The handling of EOT tokens follows train_on_eos: which defaults to turn.\n\neot_tokens:\n - \"[/INST]\"\n # - \"[/SYSTEM_PROMPT]\"\n\ndatasets:\n - path: ...\n type: chat_template\n\n # optional\n train_on_eot: turn # defaults read from train_on_eos (which defaults to turn)\n\n\n\n\n\n\nTip\n\n\n\nSee config documentation for detailed explanations of “turn”, “last”, and “all” options for training on tokens.\n\n\n\n\n\n\n\n\nNote\n\n\n\nUsing eot_tokens requires each token that exists in chat_template to be a single token in the tokenizer. Otherwise, the tokenizer will split the token and cause unexpected behavior.\nYou can add those tokens as new tokens under tokens: or (recommended) override unused added_tokens via added_tokens_overrides:. See config for more details.\n\n\n\nContinuing from the previous example, if you want to train on all EOT token trainable turns but only last EOS token, set train_on_eos: last.\n\neot_tokens:\n - \"[/INST]\"\n # ...\n\ndatasets:\n - path: ...\n type: chat_template\n\n train_on_eos: last\n train_on_eot: turn\n\n\n\n\n\n\nTip\n\n\n\nIf EOS token only appears at the end of a prompt, train_on_eos: last is equivalent to train_on_eos: turn. Therefore, generally, you can leave them to their defaults and omit them.\n\n\n\n\n\nInstead of passing tools via the system prompt, an alternative method would be to have the tools in a separate column and loaded via chat_template to let the template dynamically build it.\n{\n \"tools\": [\n {\n \"type\": \"...\",\n \"function\": {\n \"name\": \"...\",\n \"description\": \"...\",\n \"parameters\": {\n \"type\": \"...\",\n \"properties\": {\n // ...\n },\n \"required\": [\"...\"],\n },\n },\n },\n ],\n \"messages\": [\n // ...\n {\n \"role\": \"assistant\", // call the function via assistant\n \"tool_calls\": [\n {\n \"id\": \"...\", // required only for mistral\n \"type\": \"function\",\n \"function\": {\n \"name\": \"...\",\n \"arguments\": {\n \"...\": \"...\",\n }\n }\n }\n ]\n },\n {\n \"role\": \"tool\",\n \"tool_call_id\": \"...\", // required only for mistral\n \"name\": \"...\",\n \"content\": \"...\"\n },\n ],\n}\n\n\n\n\n\n\nNote\n\n\n\nTools need to follow JSON schema.\n\n\nExample config for Llama4:\nchat_template: llama4\ndatasets:\n - path: Nanobit/text-tools-2k-test\n type: chat_template\n # field_tools: tools # default is `tools`\n\n\n\n\n\n\nTip\n\n\n\nLook into the chat_template you are using to see if it supports tools and what the expected role is for the tool answer. In the example above, the tool answer is expected to be in the tool or ipython role for llama4 template.\n\n\n\n\n\n(Advanced) Using fine-grained control over tokens and turns to train in a conversation\nFor a data sample that looks like:\n\n\ndata.jsonl\n\n{\n \"conversations\": [\n {\"from\": \"system\", \"value\": \"You are an AI assistant.\", \"train\": false},\n {\"from\": \"human\", \"value\": \"Hello\", \"train\": false},\n {\"from\": \"assistant\", \"value\": \"Hello\", \"train\": true},\n {\"from\": \"human\", \"value\": \"How are you?\", \"train\": true},\n {\n \"from\": \"assistant\",\n \"value\": \"I'm doing very well, thank you!\",\n \"train_detail\": [\n {\"begin_offset\": 0, \"end_offset\": 8, \"train\": false},\n {\"begin_offset\": 9, \"end_offset\": 18, \"train\": true},\n {\"begin_offset\": 19, \"end_offset\": 30, \"train\": false},\n ],\n },\n {\n \"from\": \"human\",\n \"value\": \"I'm doing very well, thank you!\",\n \"train\": true,\n },\n {\"from\": \"assistant\", \"value\": \"Hi there!\", \"train\": true}\n ]\n}\n\nThe configuration would look like:\ndatasets:\n - path: ...\n type: chat_template\n chat_template: tokenizer_default\n field_messages: conversations\n message_property_mappings:\n role: from\n content: value\n roles_to_train: []\n train_on_eos: turn\n message_field_training: train\n message_field_training_detail: train_detail\n\n\n\n\n\n\nTip\n\n\n\nIt is not necessary to set both message_field_training and message_field_training_detail at once.\n\n\n\n\n\n(For Qwen3 template only) Enable reasoning split, where the reasoning is split from the content and passed as a separate field into the template.\ndatasets:\n - path: ...\n type: chat_template\n chat_template: qwen3\n split_thinking: true\nFor example, a content can look like:\n{\n \"content\": \"<think>Some thinking outputs</think>Output after thinking.\"\n}\nAfter split, it will look like:\n{\n \"reasoning_content\": \"Some thinking outputs\",\n \"content\": \"Output after thinking...\"\n}", "crumbs": [ "Dataset Formats", "Conversation" @@ -3326,14 +3326,14 @@ "href": "docs/api/utils.ctx_managers.sequence_parallel.html", "title": "utils.ctx_managers.sequence_parallel", "section": "", - "text": "utils.ctx_managers.sequence_parallel\nModule for Axolotl trainer sequence parallelism manager and utilities\n\n\n\n\n\nName\nDescription\n\n\n\n\nAllGatherWithGrad\nCustom autograd function for all-gather to preserve gradients.\n\n\nSequenceParallelContextManager\nContext manager for sequence parallelism operations.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad()\nCustom autograd function for all-gather to preserve gradients.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass for all-gather operation.\n\n\nforward\nForward pass of all-gather of data with sequence dimension.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad.backward(\n ctx,\n grad_output,\n)\nBackward pass for all-gather operation.\nExtracts the gradient slice corresponding to this rank’s original input\nfrom the full gradient tensor.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\ntorch.autograd function context.\nrequired\n\n\ngrad_output\ntorch.Tensor\nGradient from subsequent layers with respect to the concatenated output tensor.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, None]\nTuple containing the gradient slice for this rank’s input tensor and None for the process group parameter which doesn’t require gradients.\n\n\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad.forward(\n ctx,\n input_tensor,\n group,\n)\nForward pass of all-gather of data with sequence dimension.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\ntorch.autograd function context.\nrequired\n\n\ninput_tensor\ntorch.Tensor\nTensor from model output with sequence dimension.\nrequired\n\n\ngroup\ndist.ProcessGroup\ntorch.distributed process group.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nTensor from gathering the input_tensor from across the process group and concatenating along the sequence dimension.\n\n\n\n\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.SequenceParallelContextManager(\n models,\n context_parallel_size,\n gradient_accumulation_steps,\n ring_attn_func,\n heads_k_stride,\n gather_outputs,\n)\nContext manager for sequence parallelism operations.\nThis class provides a context that will automatically apply sequence parallelism\nduring model forward passes using a pre-forward hook, and gather outputs from\nacross the sequence parallelism group using a post-forward hook.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodels\nlist[nn.Module]\nList of models to apply sequence parallelism to pre- and post- forward hooks.\nrequired\n\n\ncontext_parallel_size\nint\nNumber of processes to split sequences over.\nrequired\n\n\ngradient_accumulation_steps\nint\nNumber of steps to accumulate gradients over.\nrequired\n\n\nring_attn_func\nRingAttnFunc\nWhich ring attention function to use. Currently unused.\nrequired\n\n\nheads_k_stride\nint | None\nSequence parallelism K head stride size. Passed through to varlen_llama3 ring_flash_attn implementation.\nrequired\n\n\ngather_outputs\nbool\nWhether to gather outputs after model forward pass across the sequence parallel group.\nrequired\n\n\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\napply_sequence_parallelism\nApply sequence parallelism slicing to a batch.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.apply_sequence_parallelism(\n batch,\n local_rank,\n local_world_size,\n gradient_accumulation_steps,\n ring_attn_func,\n)\nApply sequence parallelism slicing to a batch.\nSpecial handling is implemented for integer logits_to_keep, which indicates\nto only keep the last N tokens in the sequence during generation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nbatch\ndict[str, torch.Tensor]\nBatch dictionary (e.g., input_ids, attention_mask, etc.).\nrequired\n\n\nlocal_rank\nint\nLocal rank in the sequence parallel group.\nrequired\n\n\nlocal_world_size\nint\nWorld size of the sequence parallel group.\nrequired\n\n\ngradient_accumulation_steps\nint\nNumber of steps to accumulate gradients over.\nrequired\n\n\nring_attn_func\nRingAttnFunc\nWhich ring attention function to use. Currently unused, but related to above TODO.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[dict[str, torch.Tensor], int, int]\ntuple of: - Batch dictionary with sliced tensors. - The original sequence length before padding. - The number of padding tokens added." + "text": "utils.ctx_managers.sequence_parallel\nModule for Axolotl trainer sequence parallelism manager and utilities\n\n\n\n\n\nName\nDescription\n\n\n\n\nAllGatherWithGrad\nCustom autograd function for all-gather to preserve gradients.\n\n\nSequenceParallelContextManager\nContext manager for sequence parallelism operations.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad()\nCustom autograd function for all-gather to preserve gradients.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass for all-gather operation.\n\n\nforward\nForward pass of all-gather of data with sequence dimension.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad.backward(\n ctx,\n grad_output,\n)\nBackward pass for all-gather operation.\nExtracts the gradient slice corresponding to this rank’s original input\nfrom the full gradient tensor.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\ntorch.autograd function context.\nrequired\n\n\ngrad_output\ntorch.Tensor\nGradient from subsequent layers with respect to the concatenated output tensor.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, None]\nTuple containing the gradient slice for this rank’s input tensor and None for the process group parameter which doesn’t require gradients.\n\n\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad.forward(\n ctx,\n input_tensor,\n group,\n)\nForward pass of all-gather of data with sequence dimension.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\ntorch.autograd function context.\nrequired\n\n\ninput_tensor\ntorch.Tensor\nTensor from model output with sequence dimension.\nrequired\n\n\ngroup\ndist.ProcessGroup\ntorch.distributed process group.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nTensor from gathering the input_tensor from across the process group and concatenating along the sequence dimension.\n\n\n\n\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.SequenceParallelContextManager(\n models,\n context_parallel_size,\n gradient_accumulation_steps,\n ring_attn_func,\n heads_k_stride,\n gather_outputs,\n device_mesh=None,\n)\nContext manager for sequence parallelism operations.\nThis class provides a context that will automatically apply sequence parallelism\nduring model forward passes using a pre-forward hook, and gather outputs from\nacross the sequence parallelism group using a post-forward hook.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodels\nlist[nn.Module]\nList of models to apply sequence parallelism to pre- and post- forward hooks.\nrequired\n\n\ncontext_parallel_size\nint\nNumber of processes to split sequences over.\nrequired\n\n\ngradient_accumulation_steps\nint\nNumber of steps to accumulate gradients over.\nrequired\n\n\nring_attn_func\nRingAttnFunc\nWhich ring attention function to use. Currently unused.\nrequired\n\n\nheads_k_stride\nint | None\nSequence parallelism K head stride size. Passed through to varlen_llama3 ring_flash_attn implementation.\nrequired\n\n\ngather_outputs\nbool\nWhether to gather outputs after model forward pass across the sequence parallel group.\nrequired\n\n\n\n\n\n\n\n\n\n\n\nName\nDescription\n\n\n\n\napply_sequence_parallelism\nApply sequence parallelism slicing to a batch.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.apply_sequence_parallelism(\n batch,\n local_rank,\n local_world_size,\n gradient_accumulation_steps,\n ring_attn_func,\n)\nApply sequence parallelism slicing to a batch.\nSpecial handling is implemented for integer logits_to_keep, which indicates\nto only keep the last N tokens in the sequence during generation.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nbatch\ndict[str, torch.Tensor]\nBatch dictionary (e.g., input_ids, attention_mask, etc.).\nrequired\n\n\nlocal_rank\nint\nLocal rank in the sequence parallel group.\nrequired\n\n\nlocal_world_size\nint\nWorld size of the sequence parallel group.\nrequired\n\n\ngradient_accumulation_steps\nint\nNumber of steps to accumulate gradients over.\nrequired\n\n\nring_attn_func\nRingAttnFunc\nWhich ring attention function to use. Currently unused, but related to above TODO.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[dict[str, torch.Tensor], int, int]\ntuple of: - Batch dictionary with sliced tensors. - The original sequence length before padding. - The number of padding tokens added." }, { "objectID": "docs/api/utils.ctx_managers.sequence_parallel.html#classes", "href": "docs/api/utils.ctx_managers.sequence_parallel.html#classes", "title": "utils.ctx_managers.sequence_parallel", "section": "", - "text": "Name\nDescription\n\n\n\n\nAllGatherWithGrad\nCustom autograd function for all-gather to preserve gradients.\n\n\nSequenceParallelContextManager\nContext manager for sequence parallelism operations.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad()\nCustom autograd function for all-gather to preserve gradients.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass for all-gather operation.\n\n\nforward\nForward pass of all-gather of data with sequence dimension.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad.backward(\n ctx,\n grad_output,\n)\nBackward pass for all-gather operation.\nExtracts the gradient slice corresponding to this rank’s original input\nfrom the full gradient tensor.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\ntorch.autograd function context.\nrequired\n\n\ngrad_output\ntorch.Tensor\nGradient from subsequent layers with respect to the concatenated output tensor.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, None]\nTuple containing the gradient slice for this rank’s input tensor and None for the process group parameter which doesn’t require gradients.\n\n\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad.forward(\n ctx,\n input_tensor,\n group,\n)\nForward pass of all-gather of data with sequence dimension.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\ntorch.autograd function context.\nrequired\n\n\ninput_tensor\ntorch.Tensor\nTensor from model output with sequence dimension.\nrequired\n\n\ngroup\ndist.ProcessGroup\ntorch.distributed process group.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nTensor from gathering the input_tensor from across the process group and concatenating along the sequence dimension.\n\n\n\n\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.SequenceParallelContextManager(\n models,\n context_parallel_size,\n gradient_accumulation_steps,\n ring_attn_func,\n heads_k_stride,\n gather_outputs,\n)\nContext manager for sequence parallelism operations.\nThis class provides a context that will automatically apply sequence parallelism\nduring model forward passes using a pre-forward hook, and gather outputs from\nacross the sequence parallelism group using a post-forward hook.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodels\nlist[nn.Module]\nList of models to apply sequence parallelism to pre- and post- forward hooks.\nrequired\n\n\ncontext_parallel_size\nint\nNumber of processes to split sequences over.\nrequired\n\n\ngradient_accumulation_steps\nint\nNumber of steps to accumulate gradients over.\nrequired\n\n\nring_attn_func\nRingAttnFunc\nWhich ring attention function to use. Currently unused.\nrequired\n\n\nheads_k_stride\nint | None\nSequence parallelism K head stride size. Passed through to varlen_llama3 ring_flash_attn implementation.\nrequired\n\n\ngather_outputs\nbool\nWhether to gather outputs after model forward pass across the sequence parallel group.\nrequired" + "text": "Name\nDescription\n\n\n\n\nAllGatherWithGrad\nCustom autograd function for all-gather to preserve gradients.\n\n\nSequenceParallelContextManager\nContext manager for sequence parallelism operations.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad()\nCustom autograd function for all-gather to preserve gradients.\n\n\n\n\n\nName\nDescription\n\n\n\n\nbackward\nBackward pass for all-gather operation.\n\n\nforward\nForward pass of all-gather of data with sequence dimension.\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad.backward(\n ctx,\n grad_output,\n)\nBackward pass for all-gather operation.\nExtracts the gradient slice corresponding to this rank’s original input\nfrom the full gradient tensor.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\ntorch.autograd function context.\nrequired\n\n\ngrad_output\ntorch.Tensor\nGradient from subsequent layers with respect to the concatenated output tensor.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntuple[torch.Tensor, None]\nTuple containing the gradient slice for this rank’s input tensor and None for the process group parameter which doesn’t require gradients.\n\n\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.AllGatherWithGrad.forward(\n ctx,\n input_tensor,\n group,\n)\nForward pass of all-gather of data with sequence dimension.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nctx\ntorch.autograd.function.FunctionCtx\ntorch.autograd function context.\nrequired\n\n\ninput_tensor\ntorch.Tensor\nTensor from model output with sequence dimension.\nrequired\n\n\ngroup\ndist.ProcessGroup\ntorch.distributed process group.\nrequired\n\n\n\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\n\n\n\n\n\ntorch.Tensor\nTensor from gathering the input_tensor from across the process group and concatenating along the sequence dimension.\n\n\n\n\n\n\n\n\n\nutils.ctx_managers.sequence_parallel.SequenceParallelContextManager(\n models,\n context_parallel_size,\n gradient_accumulation_steps,\n ring_attn_func,\n heads_k_stride,\n gather_outputs,\n device_mesh=None,\n)\nContext manager for sequence parallelism operations.\nThis class provides a context that will automatically apply sequence parallelism\nduring model forward passes using a pre-forward hook, and gather outputs from\nacross the sequence parallelism group using a post-forward hook.\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\nmodels\nlist[nn.Module]\nList of models to apply sequence parallelism to pre- and post- forward hooks.\nrequired\n\n\ncontext_parallel_size\nint\nNumber of processes to split sequences over.\nrequired\n\n\ngradient_accumulation_steps\nint\nNumber of steps to accumulate gradients over.\nrequired\n\n\nring_attn_func\nRingAttnFunc\nWhich ring attention function to use. Currently unused.\nrequired\n\n\nheads_k_stride\nint | None\nSequence parallelism K head stride size. Passed through to varlen_llama3 ring_flash_attn implementation.\nrequired\n\n\ngather_outputs\nbool\nWhether to gather outputs after model forward pass across the sequence parallel group.\nrequired" }, { "objectID": "docs/api/utils.ctx_managers.sequence_parallel.html#functions", @@ -3585,14 +3585,14 @@ "href": "docs/api/utils.collators.batching.html", "title": "utils.collators.batching", "section": "", - "text": "utils.collators.batching\nData collators for axolotl to pad labels and position_ids for packed sequences\n\n\n\n\n\nName\nDescription\n\n\n\n\nBatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\nDataCollatorForSeq2Seq\nData collator that will dynamically pad the inputs received, as well as the labels and position_ids\n\n\nPretrainingBatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\nV2BatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\n\n\n\nutils.collators.batching.BatchSamplerDataCollatorForSeq2Seq(\n tokenizer,\n model=None,\n padding=True,\n max_length=None,\n pad_to_multiple_of=None,\n label_pad_token_id=-100,\n position_pad_token_id=0,\n return_tensors='pt',\n)\nCollator for multipack specific to the using the BatchSampler\n\n\n\nutils.collators.batching.DataCollatorForSeq2Seq(\n tokenizer,\n model=None,\n padding=True,\n max_length=None,\n pad_to_multiple_of=None,\n label_pad_token_id=-100,\n position_pad_token_id=0,\n return_tensors='pt',\n)\nData collator that will dynamically pad the inputs received, as well as the labels and position_ids\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntokenizer\n[PreTrainedTokenizer] or [PreTrainedTokenizerFast]\nThe tokenizer used for encoding the data.\nrequired\n\n\nmodel\n[PreTrainedModel]\nThe model that is being trained. If set and has the prepare_decoder_input_ids_from_labels, use it to prepare the decoder_input_ids This is useful when using label_smoothing to avoid calculating loss twice.\nNone\n\n\npadding\nbool, str or [~utils.PaddingStrategy], optional, defaults to True\nSelect a strategy to pad the returned sequences (according to the model’s padding side and padding index) among: - True or 'longest' (default): Pad to the longest sequence in the batch (or no padding if only a single sequence is provided). - 'max_length': Pad to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided. - False or 'do_not_pad': No padding (i.e., can output a batch with sequences of different lengths).\nTrue\n\n\nmax_length\nint, optional\nMaximum length of the returned list and optionally padding length (see above).\nNone\n\n\npad_to_multiple_of\nint, optional\nIf set will pad the sequence to a multiple of the provided value. This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).\nNone\n\n\nlabel_pad_token_id\nint, optional, defaults to -100\nThe id to use when padding the labels (-100 will be automatically ignored by PyTorch loss functions).\n-100\n\n\nreturn_tensors\nstr\nThe type of Tensor to return. Allowable values are “np”, “pt” and “tf”.\n'pt'\n\n\n\n\n\n\n\nutils.collators.batching.PretrainingBatchSamplerDataCollatorForSeq2Seq(\n *args,\n multipack_attn=True,\n **kwargs,\n)\nCollator for multipack specific to the using the BatchSampler\n\n\n\nutils.collators.batching.V2BatchSamplerDataCollatorForSeq2Seq(\n tokenizer,\n model=None,\n padding=True,\n max_length=None,\n pad_to_multiple_of=None,\n label_pad_token_id=-100,\n position_pad_token_id=0,\n return_tensors='pt',\n)\nCollator for multipack specific to the using the BatchSampler" + "text": "utils.collators.batching\nData collators for axolotl to pad labels and position_ids for packed sequences\n\n\n\n\n\nName\nDescription\n\n\n\n\nBatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\nDataCollatorForSeq2Seq\nData collator that will dynamically pad the inputs received, as well as the labels and position_ids\n\n\nPretrainingBatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\nV2BatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\n\n\n\nutils.collators.batching.BatchSamplerDataCollatorForSeq2Seq(\n tokenizer,\n model=None,\n padding=True,\n max_length=None,\n pad_to_multiple_of=None,\n label_pad_token_id=-100,\n position_pad_token_id=0,\n return_tensors='pt',\n)\nCollator for multipack specific to the using the BatchSampler\n\n\n\nutils.collators.batching.DataCollatorForSeq2Seq(\n tokenizer,\n model=None,\n padding=True,\n max_length=None,\n pad_to_multiple_of=None,\n label_pad_token_id=-100,\n position_pad_token_id=0,\n return_tensors='pt',\n)\nData collator that will dynamically pad the inputs received, as well as the labels and position_ids\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntokenizer\n[PreTrainedTokenizer] or [PreTrainedTokenizerFast]\nThe tokenizer used for encoding the data.\nrequired\n\n\nmodel\n[PreTrainedModel]\nThe model that is being trained. If set and has the prepare_decoder_input_ids_from_labels, use it to prepare the decoder_input_ids This is useful when using label_smoothing to avoid calculating loss twice.\nNone\n\n\npadding\nbool, str or [~utils.PaddingStrategy], optional, defaults to True\nSelect a strategy to pad the returned sequences (according to the model’s padding side and padding index) among: - True or 'longest' (default): Pad to the longest sequence in the batch (or no padding if only a single sequence is provided). - 'max_length': Pad to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided. - False or 'do_not_pad': No padding (i.e., can output a batch with sequences of different lengths).\nTrue\n\n\nmax_length\nint, optional\nMaximum length of the returned list and optionally padding length (see above).\nNone\n\n\npad_to_multiple_of\nint, optional\nIf set will pad the sequence to a multiple of the provided value. This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).\nNone\n\n\nlabel_pad_token_id\nint, optional, defaults to -100\nThe id to use when padding the labels (-100 will be automatically ignored by PyTorch loss functions).\n-100\n\n\nreturn_tensors\nstr\nThe type of Tensor to return. Allowable values are “np”, “pt” and “tf”.\n'pt'\n\n\n\n\n\n\n\nutils.collators.batching.PretrainingBatchSamplerDataCollatorForSeq2Seq(\n *args,\n multipack_attn=True,\n **kwargs,\n)\nCollator for multipack specific to the using the BatchSampler\n\n\n\nutils.collators.batching.V2BatchSamplerDataCollatorForSeq2Seq(\n tokenizer,\n model=None,\n padding=True,\n max_length=None,\n pad_to_multiple_of=None,\n label_pad_token_id=-100,\n position_pad_token_id=0,\n return_tensors='pt',\n squash_position_ids=False,\n)\nCollator for multipack specific to the using the BatchSampler" }, { "objectID": "docs/api/utils.collators.batching.html#classes", "href": "docs/api/utils.collators.batching.html#classes", "title": "utils.collators.batching", "section": "", - "text": "Name\nDescription\n\n\n\n\nBatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\nDataCollatorForSeq2Seq\nData collator that will dynamically pad the inputs received, as well as the labels and position_ids\n\n\nPretrainingBatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\nV2BatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\n\n\n\nutils.collators.batching.BatchSamplerDataCollatorForSeq2Seq(\n tokenizer,\n model=None,\n padding=True,\n max_length=None,\n pad_to_multiple_of=None,\n label_pad_token_id=-100,\n position_pad_token_id=0,\n return_tensors='pt',\n)\nCollator for multipack specific to the using the BatchSampler\n\n\n\nutils.collators.batching.DataCollatorForSeq2Seq(\n tokenizer,\n model=None,\n padding=True,\n max_length=None,\n pad_to_multiple_of=None,\n label_pad_token_id=-100,\n position_pad_token_id=0,\n return_tensors='pt',\n)\nData collator that will dynamically pad the inputs received, as well as the labels and position_ids\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntokenizer\n[PreTrainedTokenizer] or [PreTrainedTokenizerFast]\nThe tokenizer used for encoding the data.\nrequired\n\n\nmodel\n[PreTrainedModel]\nThe model that is being trained. If set and has the prepare_decoder_input_ids_from_labels, use it to prepare the decoder_input_ids This is useful when using label_smoothing to avoid calculating loss twice.\nNone\n\n\npadding\nbool, str or [~utils.PaddingStrategy], optional, defaults to True\nSelect a strategy to pad the returned sequences (according to the model’s padding side and padding index) among: - True or 'longest' (default): Pad to the longest sequence in the batch (or no padding if only a single sequence is provided). - 'max_length': Pad to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided. - False or 'do_not_pad': No padding (i.e., can output a batch with sequences of different lengths).\nTrue\n\n\nmax_length\nint, optional\nMaximum length of the returned list and optionally padding length (see above).\nNone\n\n\npad_to_multiple_of\nint, optional\nIf set will pad the sequence to a multiple of the provided value. This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).\nNone\n\n\nlabel_pad_token_id\nint, optional, defaults to -100\nThe id to use when padding the labels (-100 will be automatically ignored by PyTorch loss functions).\n-100\n\n\nreturn_tensors\nstr\nThe type of Tensor to return. Allowable values are “np”, “pt” and “tf”.\n'pt'\n\n\n\n\n\n\n\nutils.collators.batching.PretrainingBatchSamplerDataCollatorForSeq2Seq(\n *args,\n multipack_attn=True,\n **kwargs,\n)\nCollator for multipack specific to the using the BatchSampler\n\n\n\nutils.collators.batching.V2BatchSamplerDataCollatorForSeq2Seq(\n tokenizer,\n model=None,\n padding=True,\n max_length=None,\n pad_to_multiple_of=None,\n label_pad_token_id=-100,\n position_pad_token_id=0,\n return_tensors='pt',\n)\nCollator for multipack specific to the using the BatchSampler" + "text": "Name\nDescription\n\n\n\n\nBatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\nDataCollatorForSeq2Seq\nData collator that will dynamically pad the inputs received, as well as the labels and position_ids\n\n\nPretrainingBatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\nV2BatchSamplerDataCollatorForSeq2Seq\nCollator for multipack specific to the using the BatchSampler\n\n\n\n\n\nutils.collators.batching.BatchSamplerDataCollatorForSeq2Seq(\n tokenizer,\n model=None,\n padding=True,\n max_length=None,\n pad_to_multiple_of=None,\n label_pad_token_id=-100,\n position_pad_token_id=0,\n return_tensors='pt',\n)\nCollator for multipack specific to the using the BatchSampler\n\n\n\nutils.collators.batching.DataCollatorForSeq2Seq(\n tokenizer,\n model=None,\n padding=True,\n max_length=None,\n pad_to_multiple_of=None,\n label_pad_token_id=-100,\n position_pad_token_id=0,\n return_tensors='pt',\n)\nData collator that will dynamically pad the inputs received, as well as the labels and position_ids\n\n\n\n\n\n\n\n\n\n\n\nName\nType\nDescription\nDefault\n\n\n\n\ntokenizer\n[PreTrainedTokenizer] or [PreTrainedTokenizerFast]\nThe tokenizer used for encoding the data.\nrequired\n\n\nmodel\n[PreTrainedModel]\nThe model that is being trained. If set and has the prepare_decoder_input_ids_from_labels, use it to prepare the decoder_input_ids This is useful when using label_smoothing to avoid calculating loss twice.\nNone\n\n\npadding\nbool, str or [~utils.PaddingStrategy], optional, defaults to True\nSelect a strategy to pad the returned sequences (according to the model’s padding side and padding index) among: - True or 'longest' (default): Pad to the longest sequence in the batch (or no padding if only a single sequence is provided). - 'max_length': Pad to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided. - False or 'do_not_pad': No padding (i.e., can output a batch with sequences of different lengths).\nTrue\n\n\nmax_length\nint, optional\nMaximum length of the returned list and optionally padding length (see above).\nNone\n\n\npad_to_multiple_of\nint, optional\nIf set will pad the sequence to a multiple of the provided value. This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta).\nNone\n\n\nlabel_pad_token_id\nint, optional, defaults to -100\nThe id to use when padding the labels (-100 will be automatically ignored by PyTorch loss functions).\n-100\n\n\nreturn_tensors\nstr\nThe type of Tensor to return. Allowable values are “np”, “pt” and “tf”.\n'pt'\n\n\n\n\n\n\n\nutils.collators.batching.PretrainingBatchSamplerDataCollatorForSeq2Seq(\n *args,\n multipack_attn=True,\n **kwargs,\n)\nCollator for multipack specific to the using the BatchSampler\n\n\n\nutils.collators.batching.V2BatchSamplerDataCollatorForSeq2Seq(\n tokenizer,\n model=None,\n padding=True,\n max_length=None,\n pad_to_multiple_of=None,\n label_pad_token_id=-100,\n position_pad_token_id=0,\n return_tensors='pt',\n squash_position_ids=False,\n)\nCollator for multipack specific to the using the BatchSampler" }, { "objectID": "docs/api/prompt_strategies.input_output.html", diff --git a/sitemap.xml b/sitemap.xml index aec75b6ec..e96c9704e 100644 --- a/sitemap.xml +++ b/sitemap.xml @@ -2,794 +2,794 @@ https://docs.axolotl.ai/TODO.html - 2025-08-07T09:34:35.110Z + 2025-08-08T01:22:23.878Z https://docs.axolotl.ai/index.html - 2025-08-07T09:34:35.131Z + 2025-08-08T01:22:23.899Z https://docs.axolotl.ai/docs/debugging.html - 2025-08-07T09:34:35.112Z + 2025-08-08T01:22:23.880Z https://docs.axolotl.ai/docs/amd_hpc.html - 2025-08-07T09:34:35.111Z + 2025-08-08T01:22:23.879Z https://docs.axolotl.ai/docs/api/utils.callbacks.mlflow_.html - 2025-08-07T09:37:39.781Z + 2025-08-08T01:25:47.962Z https://docs.axolotl.ai/docs/api/monkeypatch.llama_expand_mask.html - 2025-08-07T09:37:39.210Z + 2025-08-08T01:25:47.374Z https://docs.axolotl.ai/docs/api/loaders.patch_manager.html - 2025-08-07T09:37:38.850Z + 2025-08-08T01:25:47.011Z https://docs.axolotl.ai/docs/api/core.chat.format.llama3x.html - 2025-08-07T09:37:38.533Z + 2025-08-08T01:25:46.690Z https://docs.axolotl.ai/docs/api/cli.train.html - 2025-08-07T09:37:38.590Z + 2025-08-08T01:25:46.747Z https://docs.axolotl.ai/docs/api/utils.callbacks.perplexity.html - 2025-08-07T09:37:39.772Z + 2025-08-08T01:25:47.948Z https://docs.axolotl.ai/docs/api/core.chat.messages.html - 2025-08-07T09:37:38.531Z + 2025-08-08T01:25:46.687Z https://docs.axolotl.ai/docs/api/utils.callbacks.lisa.html - 2025-08-07T09:37:39.778Z + 2025-08-08T01:25:47.958Z https://docs.axolotl.ai/docs/api/cli.merge_sharded_fsdp_weights.html - 2025-08-07T09:37:38.684Z + 2025-08-08T01:25:46.842Z https://docs.axolotl.ai/docs/api/monkeypatch.mixtral.html - 2025-08-07T09:37:39.268Z + 2025-08-08T01:25:47.434Z https://docs.axolotl.ai/docs/api/utils.chat_templates.html - 2025-08-07T09:37:39.305Z + 2025-08-08T01:25:47.472Z https://docs.axolotl.ai/docs/api/core.chat.format.shared.html - 2025-08-07T09:37:38.535Z + 2025-08-08T01:25:46.692Z https://docs.axolotl.ai/docs/api/core.trainers.mixins.optimizer.html - 2025-08-07T09:37:38.857Z + 2025-08-08T01:25:47.018Z https://docs.axolotl.ai/docs/api/utils.collators.mamba.html - 2025-08-07T09:37:39.721Z + 2025-08-08T01:25:47.897Z https://docs.axolotl.ai/docs/api/logging_config.html - 2025-08-07T09:37:38.479Z + 2025-08-08T01:25:46.636Z https://docs.axolotl.ai/docs/api/utils.collators.mm_chat.html - 2025-08-07T09:37:39.726Z + 2025-08-08T01:25:47.901Z https://docs.axolotl.ai/docs/api/prompt_strategies.completion.html - 2025-08-07T09:37:38.980Z + 2025-08-08T01:25:47.141Z https://docs.axolotl.ai/docs/api/kernels.utils.html - 2025-08-07T09:37:39.195Z + 2025-08-08T01:25:47.359Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.chat_template.html - 2025-08-07T09:37:39.013Z + 2025-08-08T01:25:47.174Z https://docs.axolotl.ai/docs/api/kernels.swiglu.html - 2025-08-07T09:37:39.186Z + 2025-08-08T01:25:47.350Z https://docs.axolotl.ai/docs/api/common.const.html - 2025-08-07T09:37:39.682Z + 2025-08-08T01:25:47.857Z https://docs.axolotl.ai/docs/api/cli.cloud.base.html - 2025-08-07T09:37:38.707Z + 2025-08-08T01:25:46.866Z https://docs.axolotl.ai/docs/api/prompt_strategies.orpo.chat_template.html - 2025-08-07T09:37:39.075Z + 2025-08-08T01:25:47.237Z https://docs.axolotl.ai/docs/api/core.builders.rl.html - 2025-08-07T09:37:38.495Z + 2025-08-08T01:25:46.652Z https://docs.axolotl.ai/docs/api/utils.dict.html - 2025-08-07T09:37:39.396Z + 2025-08-08T01:25:47.565Z https://docs.axolotl.ai/docs/api/utils.schemas.integrations.html - 2025-08-07T09:37:39.506Z + 2025-08-08T01:25:47.678Z https://docs.axolotl.ai/docs/api/core.trainers.utils.html - 2025-08-07T09:37:38.815Z + 2025-08-08T01:25:46.976Z https://docs.axolotl.ai/docs/api/monkeypatch.trainer_fsdp_optim.html - 2025-08-07T09:37:39.258Z + 2025-08-08T01:25:47.423Z https://docs.axolotl.ai/docs/api/cli.evaluate.html - 2025-08-07T09:37:38.598Z + 2025-08-08T01:25:46.755Z https://docs.axolotl.ai/docs/api/core.builders.causal.html - 2025-08-07T09:37:38.490Z + 2025-08-08T01:25:46.647Z https://docs.axolotl.ai/docs/api/monkeypatch.multipack.html - 2025-08-07T09:37:39.205Z + 2025-08-08T01:25:47.369Z https://docs.axolotl.ai/docs/api/monkeypatch.llama_patch_multipack.html - 2025-08-07T09:37:39.249Z + 2025-08-08T01:25:47.414Z https://docs.axolotl.ai/docs/api/cli.delinearize_llama4.html - 2025-08-07T09:37:38.649Z + 2025-08-08T01:25:46.807Z https://docs.axolotl.ai/docs/api/utils.schemas.trl.html - 2025-08-07T09:37:39.489Z + 2025-08-08T01:25:47.660Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.zephyr.html - 2025-08-07T09:37:39.035Z + 2025-08-08T01:25:47.196Z https://docs.axolotl.ai/docs/api/integrations.kd.trainer.html - 2025-08-07T09:37:39.669Z + 2025-08-08T01:25:47.844Z https://docs.axolotl.ai/docs/api/monkeypatch.gradient_checkpointing.offload_disk.html - 2025-08-07T09:37:39.297Z + 2025-08-08T01:25:47.463Z https://docs.axolotl.ai/docs/api/utils.optimizers.adopt.html - 2025-08-07T09:37:39.403Z + 2025-08-08T01:25:47.572Z https://docs.axolotl.ai/docs/api/monkeypatch.data.batch_dataset_fetcher.html - 2025-08-07T09:37:39.267Z + 2025-08-08T01:25:47.432Z https://docs.axolotl.ai/docs/api/cli.cloud.modal_.html - 2025-08-07T09:37:38.713Z + 2025-08-08T01:25:46.872Z https://docs.axolotl.ai/docs/api/prompt_strategies.alpaca_chat.html - 2025-08-07T09:37:38.940Z + 2025-08-08T01:25:47.100Z https://docs.axolotl.ai/docs/api/utils.freeze.html - 2025-08-07T09:37:39.327Z + 2025-08-08T01:25:47.493Z https://docs.axolotl.ai/docs/api/prompt_strategies.bradley_terry.llama3.html - 2025-08-07T09:37:39.079Z + 2025-08-08T01:25:47.241Z https://docs.axolotl.ai/docs/api/integrations.base.html - 2025-08-07T09:37:39.658Z + 2025-08-08T01:25:47.832Z https://docs.axolotl.ai/docs/api/monkeypatch.unsloth_.html - 2025-08-07T09:37:39.265Z + 2025-08-08T01:25:47.431Z https://docs.axolotl.ai/docs/api/prompt_strategies.kto.chatml.html - 2025-08-07T09:37:39.054Z + 2025-08-08T01:25:47.215Z https://docs.axolotl.ai/docs/api/cli.main.html - 2025-08-07T09:37:38.582Z + 2025-08-08T01:25:46.739Z https://docs.axolotl.ai/docs/api/common.datasets.html - 2025-08-07T09:37:39.697Z + 2025-08-08T01:25:47.872Z https://docs.axolotl.ai/docs/api/train.html - 2025-08-07T09:37:38.394Z + 2025-08-08T01:25:46.550Z https://docs.axolotl.ai/docs/api/utils.trainer.html - 2025-08-07T09:37:39.344Z + 2025-08-08T01:25:47.510Z https://docs.axolotl.ai/docs/api/prompt_strategies.llama2_chat.html - 2025-08-07T09:37:38.974Z + 2025-08-08T01:25:47.135Z https://docs.axolotl.ai/docs/api/index.html - 2025-08-07T09:37:38.333Z + 2025-08-08T01:25:46.488Z https://docs.axolotl.ai/docs/api/prompt_strategies.chat_template.html - 2025-08-07T09:37:38.927Z + 2025-08-08T01:25:47.087Z https://docs.axolotl.ai/docs/api/core.training_args.html - 2025-08-07T09:37:38.508Z + 2025-08-08T01:25:46.665Z https://docs.axolotl.ai/docs/api/kernels.quantize.html - 2025-08-07T09:37:39.193Z + 2025-08-08T01:25:47.358Z https://docs.axolotl.ai/docs/api/convert.html - 2025-08-07T09:37:38.428Z + 2025-08-08T01:25:46.585Z https://docs.axolotl.ai/docs/api/integrations.grokfast.optimizer.html - 2025-08-07T09:37:39.662Z + 2025-08-08T01:25:47.836Z https://docs.axolotl.ai/docs/api/prompt_strategies.stepwise_supervised.html - 2025-08-07T09:37:38.990Z + 2025-08-08T01:25:47.151Z https://docs.axolotl.ai/docs/api/utils.schemas.model.html - 2025-08-07T09:37:39.453Z + 2025-08-08T01:25:47.622Z https://docs.axolotl.ai/docs/api/utils.callbacks.qat.html - 2025-08-07T09:37:39.791Z + 2025-08-08T01:25:47.972Z https://docs.axolotl.ai/docs/api/loaders.constants.html - 2025-08-07T09:37:38.852Z + 2025-08-08T01:25:47.013Z https://docs.axolotl.ai/docs/api/cli.utils.sweeps.html - 2025-08-07T09:37:38.743Z + 2025-08-08T01:25:46.902Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.llama3.html - 2025-08-07T09:37:39.023Z + 2025-08-08T01:25:47.184Z https://docs.axolotl.ai/docs/api/core.datasets.transforms.chat_builder.html - 2025-08-07T09:37:38.548Z + 2025-08-08T01:25:46.705Z https://docs.axolotl.ai/docs/api/cli.utils.fetch.html - 2025-08-07T09:37:38.731Z + 2025-08-08T01:25:46.890Z https://docs.axolotl.ai/docs/api/core.trainers.mamba.html - 2025-08-07T09:37:38.784Z + 2025-08-08T01:25:46.944Z https://docs.axolotl.ai/docs/api/utils.schemas.enums.html - 2025-08-07T09:37:39.517Z + 2025-08-08T01:25:47.688Z https://docs.axolotl.ai/docs/api/utils.callbacks.profiler.html - 2025-08-07T09:37:39.776Z + 2025-08-08T01:25:47.957Z https://docs.axolotl.ai/docs/api/prompt_strategies.metharme.html - 2025-08-07T09:37:38.997Z + 2025-08-08T01:25:47.158Z https://docs.axolotl.ai/docs/api/core.trainers.trl.html - 2025-08-07T09:37:38.779Z + 2025-08-08T01:25:46.939Z https://docs.axolotl.ai/docs/api/prompt_strategies.orcamini.html - 2025-08-07T09:37:39.001Z + 2025-08-08T01:25:47.162Z https://docs.axolotl.ai/docs/api/utils.samplers.multipack.html - 2025-08-07T09:37:39.766Z + 2025-08-08T01:25:47.942Z https://docs.axolotl.ai/docs/api/utils.schedulers.html - 2025-08-07T09:37:39.371Z + 2025-08-08T01:25:47.539Z https://docs.axolotl.ai/docs/api/core.trainers.grpo.trainer.html - 2025-08-07T09:37:38.802Z + 2025-08-08T01:25:46.962Z https://docs.axolotl.ai/docs/api/prompt_tokenizers.html - 2025-08-07T09:37:38.470Z + 2025-08-08T01:25:46.627Z https://docs.axolotl.ai/docs/config-reference.html - 2025-08-07T09:37:53.302Z + 2025-08-08T01:26:01.870Z https://docs.axolotl.ai/docs/multimodal.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/mixed_precision.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/unsloth.html - 2025-08-07T09:34:35.116Z + 2025-08-08T01:22:23.884Z https://docs.axolotl.ai/docs/ray-integration.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/dataset-formats/stepwise_supervised.html - 2025-08-07T09:34:35.111Z + 2025-08-08T01:22:23.880Z https://docs.axolotl.ai/docs/dataset-formats/template_free.html - 2025-08-07T09:34:35.111Z + 2025-08-08T01:22:23.880Z https://docs.axolotl.ai/docs/dataset-formats/index.html - 2025-08-07T09:34:35.111Z + 2025-08-08T01:22:23.879Z https://docs.axolotl.ai/docs/dataset-formats/pretraining.html - 2025-08-07T09:34:35.111Z + 2025-08-08T01:22:23.880Z https://docs.axolotl.ai/docs/nd_parallelism.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/sequence_parallelism.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.884Z https://docs.axolotl.ai/docs/inference.html - 2025-08-07T09:34:35.114Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/fsdp_qlora.html - 2025-08-07T09:34:35.112Z + 2025-08-08T01:22:23.880Z https://docs.axolotl.ai/docs/multi-node.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/lora_optims.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/getting-started.html - 2025-08-07T09:34:35.112Z + 2025-08-08T01:22:23.880Z https://docs.axolotl.ai/docs/dataset_loading.html - 2025-08-07T09:34:35.111Z + 2025-08-08T01:22:23.880Z https://docs.axolotl.ai/docs/lr_groups.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/input_output.html - 2025-08-07T09:34:35.114Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/src/axolotl/integrations/LICENSE.html - 2025-08-07T09:34:35.135Z + 2025-08-08T01:22:23.903Z https://docs.axolotl.ai/src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html - 2025-08-07T09:34:35.135Z + 2025-08-08T01:22:23.904Z https://docs.axolotl.ai/docs/mac.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/optimizers.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/gradient_checkpointing.html - 2025-08-07T09:34:35.112Z + 2025-08-08T01:22:23.880Z https://docs.axolotl.ai/docs/qat.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/faq.html - 2025-08-07T09:34:35.112Z + 2025-08-08T01:22:23.880Z https://docs.axolotl.ai/docs/dataset_preprocessing.html - 2025-08-07T09:34:35.111Z + 2025-08-08T01:22:23.880Z https://docs.axolotl.ai/docs/nccl.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/cli.html - 2025-08-07T09:34:35.111Z + 2025-08-08T01:22:23.879Z https://docs.axolotl.ai/docs/torchao.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.884Z https://docs.axolotl.ai/docs/multi-gpu.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/rlhf.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.884Z https://docs.axolotl.ai/docs/dataset-formats/tokenized.html - 2025-08-07T09:34:35.111Z + 2025-08-08T01:22:23.880Z https://docs.axolotl.ai/docs/dataset-formats/conversation.html - 2025-08-07T09:34:35.111Z + 2025-08-08T01:22:23.879Z https://docs.axolotl.ai/docs/dataset-formats/inst_tune.html - 2025-08-07T09:34:35.111Z + 2025-08-08T01:22:23.880Z https://docs.axolotl.ai/docs/reward_modelling.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/docker.html - 2025-08-07T09:34:35.112Z + 2025-08-08T01:22:23.880Z https://docs.axolotl.ai/docs/installation.html - 2025-08-07T09:34:35.114Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/quantize.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/docs/custom_integrations.html - 2025-08-07T09:34:35.111Z + 2025-08-08T01:22:23.879Z https://docs.axolotl.ai/docs/batch_vs_grad.html - 2025-08-07T09:34:35.111Z + 2025-08-08T01:22:23.879Z https://docs.axolotl.ai/docs/api/cli.utils.train.html - 2025-08-07T09:37:38.753Z + 2025-08-08T01:25:46.913Z https://docs.axolotl.ai/docs/api/cli.art.html - 2025-08-07T09:37:38.620Z + 2025-08-08T01:25:46.778Z https://docs.axolotl.ai/docs/api/core.trainers.grpo.sampler.html - 2025-08-07T09:37:38.814Z + 2025-08-08T01:25:46.975Z https://docs.axolotl.ai/docs/api/loaders.model.html - 2025-08-07T09:37:38.825Z + 2025-08-08T01:25:46.986Z https://docs.axolotl.ai/docs/api/cli.preprocess.html - 2025-08-07T09:37:38.692Z + 2025-08-08T01:25:46.850Z https://docs.axolotl.ai/docs/api/cli.utils.html - 2025-08-07T09:37:38.714Z + 2025-08-08T01:25:46.873Z https://docs.axolotl.ai/docs/api/cli.inference.html - 2025-08-07T09:37:38.663Z + 2025-08-08T01:25:46.822Z https://docs.axolotl.ai/docs/api/monkeypatch.btlm_attn_hijack_flash.html - 2025-08-07T09:37:39.247Z + 2025-08-08T01:25:47.412Z https://docs.axolotl.ai/docs/api/datasets.html - 2025-08-07T09:37:38.415Z + 2025-08-08T01:25:46.571Z https://docs.axolotl.ai/docs/api/monkeypatch.transformers_fa_utils.html - 2025-08-07T09:37:39.264Z + 2025-08-08T01:25:47.429Z https://docs.axolotl.ai/docs/api/monkeypatch.llama_attn_hijack_flash.html - 2025-08-07T09:37:39.200Z + 2025-08-08T01:25:47.365Z https://docs.axolotl.ai/docs/api/monkeypatch.relora.html - 2025-08-07T09:37:39.208Z + 2025-08-08T01:25:47.373Z https://docs.axolotl.ai/docs/api/monkeypatch.stablelm_attn_hijack_flash.html - 2025-08-07T09:37:39.254Z + 2025-08-08T01:25:47.420Z https://docs.axolotl.ai/docs/api/loaders.adapter.html - 2025-08-07T09:37:38.840Z + 2025-08-08T01:25:47.001Z https://docs.axolotl.ai/docs/api/core.trainers.dpo.trainer.html - 2025-08-07T09:37:38.791Z + 2025-08-08T01:25:46.951Z https://docs.axolotl.ai/docs/api/integrations.cut_cross_entropy.args.html - 2025-08-07T09:37:39.661Z + 2025-08-08T01:25:47.835Z https://docs.axolotl.ai/docs/api/monkeypatch.utils.html - 2025-08-07T09:37:39.246Z + 2025-08-08T01:25:47.411Z https://docs.axolotl.ai/docs/api/loaders.processor.html - 2025-08-07T09:37:38.835Z + 2025-08-08T01:25:46.996Z https://docs.axolotl.ai/docs/api/cli.config.html - 2025-08-07T09:37:38.644Z + 2025-08-08T01:25:46.802Z https://docs.axolotl.ai/docs/api/integrations.liger.args.html - 2025-08-07T09:37:39.673Z + 2025-08-08T01:25:47.847Z https://docs.axolotl.ai/docs/api/loaders.tokenizer.html - 2025-08-07T09:37:38.833Z + 2025-08-08T01:25:46.994Z https://docs.axolotl.ai/docs/api/utils.schemas.config.html - 2025-08-07T09:37:39.446Z + 2025-08-08T01:25:47.615Z https://docs.axolotl.ai/docs/api/utils.ctx_managers.sequence_parallel.html - 2025-08-07T09:37:38.893Z + 2025-08-08T01:25:47.052Z https://docs.axolotl.ai/docs/api/core.trainers.mixins.scheduler.html - 2025-08-07T09:37:38.870Z + 2025-08-08T01:25:47.028Z https://docs.axolotl.ai/docs/api/core.trainers.base.html - 2025-08-07T09:37:38.764Z + 2025-08-08T01:25:46.923Z https://docs.axolotl.ai/docs/api/cli.utils.args.html - 2025-08-07T09:37:38.726Z + 2025-08-08T01:25:46.885Z https://docs.axolotl.ai/docs/api/prompt_strategies.messages.chat.html - 2025-08-07T09:37:39.011Z + 2025-08-08T01:25:47.173Z https://docs.axolotl.ai/docs/api/monkeypatch.lora_kernels.html - 2025-08-07T09:37:39.238Z + 2025-08-08T01:25:47.403Z https://docs.axolotl.ai/docs/api/kernels.lora.html - 2025-08-07T09:37:39.166Z + 2025-08-08T01:25:47.330Z https://docs.axolotl.ai/docs/api/cli.vllm_serve.html - 2025-08-07T09:37:38.703Z + 2025-08-08T01:25:46.862Z https://docs.axolotl.ai/docs/api/utils.schemas.multimodal.html - 2025-08-07T09:37:39.494Z + 2025-08-08T01:25:47.666Z https://docs.axolotl.ai/docs/api/utils.schemas.utils.html - 2025-08-07T09:37:39.522Z + 2025-08-08T01:25:47.694Z https://docs.axolotl.ai/docs/api/monkeypatch.llama_attn_hijack_xformers.html - 2025-08-07T09:37:39.202Z + 2025-08-08T01:25:47.366Z https://docs.axolotl.ai/docs/api/integrations.lm_eval.args.html - 2025-08-07T09:37:39.676Z + 2025-08-08T01:25:47.850Z https://docs.axolotl.ai/docs/api/monkeypatch.mistral_attn_hijack_flash.html - 2025-08-07T09:37:39.203Z + 2025-08-08T01:25:47.368Z https://docs.axolotl.ai/docs/api/utils.collators.core.html - 2025-08-07T09:37:39.699Z + 2025-08-08T01:25:47.874Z https://docs.axolotl.ai/docs/api/core.chat.format.chatml.html - 2025-08-07T09:37:38.532Z + 2025-08-08T01:25:46.689Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.passthrough.html - 2025-08-07T09:37:39.038Z + 2025-08-08T01:25:47.199Z https://docs.axolotl.ai/docs/api/core.datasets.chat.html - 2025-08-07T09:37:38.540Z + 2025-08-08T01:25:46.697Z https://docs.axolotl.ai/docs/api/utils.bench.html - 2025-08-07T09:37:39.320Z + 2025-08-08T01:25:47.485Z https://docs.axolotl.ai/docs/api/utils.schemas.training.html - 2025-08-07T09:37:39.459Z + 2025-08-08T01:25:47.629Z https://docs.axolotl.ai/docs/api/utils.collators.batching.html - 2025-08-07T09:37:39.718Z + 2025-08-08T01:25:47.893Z https://docs.axolotl.ai/docs/api/prompt_strategies.input_output.html - 2025-08-07T09:37:38.986Z + 2025-08-08T01:25:47.147Z https://docs.axolotl.ai/docs/api/utils.lora.html - 2025-08-07T09:37:39.310Z + 2025-08-08T01:25:47.476Z https://docs.axolotl.ai/docs/api/prompt_strategies.base.html - 2025-08-07T09:37:38.895Z + 2025-08-08T01:25:47.054Z https://docs.axolotl.ai/docs/api/prompt_strategies.alpaca_w_system.html - 2025-08-07T09:37:38.953Z + 2025-08-08T01:25:47.114Z https://docs.axolotl.ai/docs/api/utils.schemas.datasets.html - 2025-08-07T09:37:39.477Z + 2025-08-08T01:25:47.648Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.user_defined.html - 2025-08-07T09:37:39.036Z + 2025-08-08T01:25:47.198Z https://docs.axolotl.ai/docs/api/utils.schemas.peft.html - 2025-08-07T09:37:39.486Z + 2025-08-08T01:25:47.657Z https://docs.axolotl.ai/docs/api/prompt_strategies.pygmalion.html - 2025-08-07T09:37:39.007Z + 2025-08-08T01:25:47.168Z https://docs.axolotl.ai/docs/api/common.architectures.html - 2025-08-07T09:37:39.681Z + 2025-08-08T01:25:47.855Z https://docs.axolotl.ai/docs/api/monkeypatch.gradient_checkpointing.offload_cpu.html - 2025-08-07T09:37:39.272Z + 2025-08-08T01:25:47.437Z https://docs.axolotl.ai/docs/api/utils.callbacks.comet_.html - 2025-08-07T09:37:39.785Z + 2025-08-08T01:25:47.965Z https://docs.axolotl.ai/docs/api/integrations.spectrum.args.html - 2025-08-07T09:37:39.679Z + 2025-08-08T01:25:47.854Z https://docs.axolotl.ai/docs/api/cli.quantize.html - 2025-08-07T09:37:38.697Z + 2025-08-08T01:25:46.855Z https://docs.axolotl.ai/docs/api/cli.checks.html - 2025-08-07T09:37:38.627Z + 2025-08-08T01:25:46.785Z https://docs.axolotl.ai/docs/api/prompt_strategies.kto.llama3.html - 2025-08-07T09:37:39.046Z + 2025-08-08T01:25:47.207Z https://docs.axolotl.ai/docs/api/utils.model_shard_quant.html - 2025-08-07T09:37:39.316Z + 2025-08-08T01:25:47.482Z https://docs.axolotl.ai/docs/api/utils.quantization.html - 2025-08-07T09:37:39.432Z + 2025-08-08T01:25:47.601Z https://docs.axolotl.ai/docs/api/core.trainers.mixins.rng_state_loader.html - 2025-08-07T09:37:38.863Z + 2025-08-08T01:25:47.022Z https://docs.axolotl.ai/docs/api/kernels.geglu.html - 2025-08-07T09:37:39.176Z + 2025-08-08T01:25:47.340Z https://docs.axolotl.ai/docs/api/utils.data.pretraining.html - 2025-08-07T09:37:39.405Z + 2025-08-08T01:25:47.574Z https://docs.axolotl.ai/docs/api/prompt_strategies.kto.user_defined.html - 2025-08-07T09:37:39.055Z + 2025-08-08T01:25:47.217Z https://docs.axolotl.ai/docs/api/core.builders.base.html - 2025-08-07T09:37:38.486Z + 2025-08-08T01:25:46.642Z https://docs.axolotl.ai/docs/api/cli.merge_lora.html - 2025-08-07T09:37:38.672Z + 2025-08-08T01:25:46.830Z https://docs.axolotl.ai/docs/api/cli.utils.load.html - 2025-08-07T09:37:38.737Z + 2025-08-08T01:25:46.896Z https://docs.axolotl.ai/docs/api/utils.data.sft.html - 2025-08-07T09:37:39.412Z + 2025-08-08T01:25:47.581Z https://docs.axolotl.ai/docs/api/prompt_strategies.user_defined.html - 2025-08-07T09:37:38.961Z + 2025-08-08T01:25:47.122Z https://docs.axolotl.ai/docs/api/utils.tokenization.html - 2025-08-07T09:37:39.304Z + 2025-08-08T01:25:47.470Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.chatml.html - 2025-08-07T09:37:39.033Z + 2025-08-08T01:25:47.195Z https://docs.axolotl.ai/docs/api/models.mamba.modeling_mamba.html - 2025-08-07T09:37:39.698Z + 2025-08-08T01:25:47.873Z https://docs.axolotl.ai/docs/api/cli.args.html - 2025-08-07T09:37:38.617Z + 2025-08-08T01:25:46.775Z https://docs.axolotl.ai/docs/api/evaluate.html - 2025-08-07T09:37:38.404Z + 2025-08-08T01:25:46.560Z https://docs.axolotl.ai/docs/api/prompt_strategies.alpaca_instruct.html - 2025-08-07T09:37:38.942Z + 2025-08-08T01:25:47.102Z https://docs.axolotl.ai/docs/api/utils.distributed.html - 2025-08-07T09:37:39.391Z + 2025-08-08T01:25:47.559Z https://docs.axolotl.ai/docs/multipack.html - 2025-08-07T09:34:35.115Z + 2025-08-08T01:22:23.883Z https://docs.axolotl.ai/examples/colab-notebooks/colab-axolotl-example.html - 2025-08-07T09:34:35.119Z + 2025-08-08T01:22:23.888Z https://docs.axolotl.ai/FAQS.html - 2025-08-07T09:34:35.109Z + 2025-08-08T01:22:23.878Z