diff --git a/.nojekyll b/.nojekyll index c970f127e..0ebb97932 100644 --- a/.nojekyll +++ b/.nojekyll @@ -1 +1 @@ -a02c956e \ No newline at end of file +165a63b7 \ No newline at end of file diff --git a/docs/config.html b/docs/config.html index 454f3f7b3..856a9965e 100644 --- a/docs/config.html +++ b/docs/config.html @@ -652,554 +652,558 @@ pre > code.sourceCode > span > a:first-child::before { text-decoration: underlin # adding a system turn with empty content. drop_system_message: - # IMPORTANT: The following fields determine which parts of the conversation to train on. - # Priority order: message_field_training > message_field_training_detail > train_on_inputs or role in roles_to_train - # See examples at `docs/dataset-formats/conversation.qmd` - # Note: If the below 5 fields are empty, defaults to training only on the last message. - - # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss. - roles_to_train: ["assistant"] # default - # Optional[str]. 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 - # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`. - train_on_eos: turn - # Optional[str]. Which EOT (End-of-Turn) tokens to train on in the conversation. Possible values are: - # - all: train on all EOT tokens - # - turn: train on the EOT token at the end of each trainable turn - # - last: train on the last EOT token in the conversation - # If not specified, defaults to the value of train_on_eos for backward compatibility. - train_on_eot: - # 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: training - # 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: train_detail - - -# 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: true + # Optional[bool]. Whether to split the assistant turn based on a reasoning trace inside delimited tags + # defaults to False + split_thinking: + + # IMPORTANT: The following fields determine which parts of the conversation to train on. + # Priority order: message_field_training > message_field_training_detail > train_on_inputs or role in roles_to_train + # See examples at `docs/dataset-formats/conversation.qmd` + # Note: If the below 5 fields are empty, defaults to training only on the last message. + + # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss. + roles_to_train: ["assistant"] # default + # Optional[str]. 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 + # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`. + train_on_eos: turn + # Optional[str]. Which EOT (End-of-Turn) tokens to train on in the conversation. Possible values are: + # - all: train on all EOT tokens + # - turn: train on the EOT token at the end of each trainable turn + # - last: train on the last EOT token in the conversation + # If not specified, defaults to the value of train_on_eos for backward compatibility. + train_on_eot: + # 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: training + # 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: train_detail + -Deduplicates datasets and test_datasets with identical entries. -dataset_exact_deduplication: true - -# 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: - - path: /workspace/data/eval.jsonl - ds_type: json - # You need to specify a split. For "json" datasets the default split is called "train". - split: train - type: completion - data_files: - - /workspace/data/eval.jsonl - -# use RL training: 'dpo', 'ipo', 'kto', 'simpo', 'orpo', 'grpo' -rl: -rl_beta: # Optional[float]. The beta parameter for the RL training. +# 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: true + +Deduplicates datasets and test_datasets with identical entries. +dataset_exact_deduplication: true + +# 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: + - path: /workspace/data/eval.jsonl + ds_type: json + # You need to specify a split. For "json" datasets the default split is called "train". + split: train + type: completion + data_files: + - /workspace/data/eval.jsonl -# dpo -dpo_use_weighting: # Optional[bool]. Whether to perform weighting. -rpo_alpha: # Optional[float]. Weighting of NLL term in loss from RPO paper. +# use RL training: 'dpo', 'ipo', 'kto', 'simpo', 'orpo', 'grpo' +rl: +rl_beta: # Optional[float]. The beta parameter for the RL training. -# orpo -orpo_alpha: 0.1 # Parameter controlling the relative ratio loss weight in the ORPO loss. Passed to `beta` in `ORPOConfig` due to trl mapping. - -# kto -kto_desirable_weight: # Optional[float]. Factor for desirable loss term in KTO loss. -kto_undesirable_weight: # Optional[float]. Factor for undesirable loss term in KTO loss. +# dpo +dpo_use_weighting: # Optional[bool]. Whether to perform weighting. +rpo_alpha: # Optional[float]. Weighting of NLL term in loss from RPO paper. + +# orpo +orpo_alpha: 0.1 # Parameter controlling the relative ratio loss weight in the ORPO loss. Passed to `beta` in `ORPOConfig` due to trl mapping. -# simpo -cpo_alpha: 1.0 # Weight of the BC regularizer -simpo_gamma: 0.5 # Target reward margin for the SimPO loss +# kto +kto_desirable_weight: # Optional[float]. Factor for desirable loss term in KTO loss. +kto_undesirable_weight: # Optional[float]. Factor for undesirable loss term in KTO loss. -# grpo -trl: - use_vllm: # Optional[bool]. Whether to use VLLM for RL training. - vllm_server_host: # Optional[str]. Host of the vLLM server to connect to. - vllm_server_port: # Optional[int]. Port of the vLLM server to connect to. - vllm_server_timeout: # Optional[int]. Total timeout (in seconds) to wait for the vLLM server to respond. - vllm_guided_decoding_regex: # Optional[str]. Regex for vLLM guided decoding. - - beta: # Optional[float]. Beta parameter for the RL training. Same as `rl_beta`. Use - max_completion_length: # Optional[int]. Maximum length of the completion for RL training. - - reward_funcs: # Optional[list[str]]. List of reward functions to load. Paths must be importable from current dir. - reward_weights: # Optional[list[float]]. List of reward weights for the reward functions. - - num_generations: # Optional[int]. Number of generations to sample. - log_completions: # Optional[bool]. Whether to log completions. - - sync_ref_model: # Optional[bool]. Whether to sync the reference model. - ref_model_mixup_alpha: # Optional[float]. Mixup alpha for the reference model. - ref_model_sync_steps: # Optional[int]. Sync steps for the reference model. +# simpo +cpo_alpha: 1.0 # Weight of the BC regularizer +simpo_gamma: 0.5 # Target reward margin for the SimPO loss + +# grpo +trl: + use_vllm: # Optional[bool]. Whether to use VLLM for RL training. + vllm_server_host: # Optional[str]. Host of the vLLM server to connect to. + vllm_server_port: # Optional[int]. Port of the vLLM server to connect to. + vllm_server_timeout: # Optional[int]. Total timeout (in seconds) to wait for the vLLM server to respond. + vllm_guided_decoding_regex: # Optional[str]. Regex for vLLM guided decoding. + + beta: # Optional[float]. Beta parameter for the RL training. Same as `rl_beta`. Use + max_completion_length: # Optional[int]. Maximum length of the completion for RL training. + + reward_funcs: # Optional[list[str]]. List of reward functions to load. Paths must be importable from current dir. + reward_weights: # Optional[list[float]]. List of reward weights for the reward functions. + + num_generations: # Optional[int]. Number of generations to sample. + log_completions: # Optional[bool]. Whether to log completions. - -# reward modelling: `True` or `False` -reward_model: + sync_ref_model: # Optional[bool]. Whether to sync the reference model. + ref_model_mixup_alpha: # Optional[float]. Mixup alpha for the reference model. + ref_model_sync_steps: # Optional[int]. Sync steps for the reference model. -# process reward modelling: `True` or `False` -process_reward_model: - -# 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 -# Note: It is recommended to set train_on_inputs to true when using a chat template that is different from the model's default chat template. -chat_template: tokenizer_default -# custom jinja template 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: null -# Optional[List[str]]. 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: - # - "</s>" - # - "[/INST]" - # - "[/SYSTEM_PROMPT]" -# Changes the default system message -default_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml. -# 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: data/last_run_prepared -# Push prepared dataset to hub -push_dataset_to_hub: # Optional[str] repo_org/repo_name -# The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()` -# if not set. -dataset_processes: # defaults to os.cpu_count() if not set -# Keep dataset in memory while preprocessing -# Only needed if cached dataset is taking too much storage -dataset_keep_in_memory: -# push checkpoints to hub -hub_model_id: # private repo path to push finetuned model -# how to push checkpoints to hub -# https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy -hub_strategy: -# 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: # boolean -# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval. -val_set_size: 0.04 -# Num shards for whole dataset -dataset_shard_num: -# Index of shard to use for whole dataset -dataset_shard_idx: - -# 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: 2048 -# Pad inputs so each step uses constant sized buffers -# This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently -pad_to_sequence_len: -# Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true' -sample_packing: -# Set to 'false' if getting errors during eval with sample_packing on. -eval_sample_packing: -# 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: -total_num_tokens: -# Increasing the following values helps with packing, but usually only slightly (<%1.) -# The number of samples packed at a time. -sample_packing_group_size: 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: 200 -sample_pack_sequentially: # Optional[bool]. Whether to pack samples sequentially. - -# whether to concatenate samples during pretraining -pretraining_sample_concatenation: - -curriculum_sampling: # Optional[bool]. Whether to use sequential sampling for curriculum learning - -# Use batch flattening for speedups when not using sample_packing -batch_flattening: - -# Passed through to transformers when loading the model when launched without accelerate -# Use `sequential` when training w/ model parallelism to limit memory -device_map: -# Defines the max memory usage per gpu on the system. Passed through to transformers when loading the model. -max_memory: - -# If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model -adapter: lora -# 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: - -# LoRA hyperparameters -# For more details about the following options, see: -# https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2 -lora_r: 8 -lora_alpha: 16 -lora_dropout: 0.05 -lora_target_modules: - - q_proj - - v_proj -# - k_proj -# - o_proj -# - gate_proj -# - down_proj -# - up_proj -lora_target_linear: # If true, will target all linear modules - -# List[int] | int. # The layer indices to transform, otherwise, apply to all layers -# https://huggingface.co/docs/peft/v0.15.0/en/package_reference/lora#peft.LoraConfig.layers_to_transform -peft_layers_to_transform: + +# reward modelling: `True` or `False` +reward_model: + +# process reward modelling: `True` or `False` +process_reward_model: + +# 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 +# Note: It is recommended to set train_on_inputs to true when using a chat template that is different from the model's default chat template. +chat_template: tokenizer_default +# custom jinja template 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: null +# Optional[List[str]]. 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: + # - "</s>" + # - "[/INST]" + # - "[/SYSTEM_PROMPT]" +# Changes the default system message +default_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml. +# 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: data/last_run_prepared +# Push prepared dataset to hub +push_dataset_to_hub: # Optional[str] repo_org/repo_name +# The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()` +# if not set. +dataset_processes: # defaults to os.cpu_count() if not set +# Keep dataset in memory while preprocessing +# Only needed if cached dataset is taking too much storage +dataset_keep_in_memory: +# push checkpoints to hub +hub_model_id: # private repo path to push finetuned model +# how to push checkpoints to hub +# https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy +hub_strategy: +# 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: # boolean +# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval. +val_set_size: 0.04 +# Num shards for whole dataset +dataset_shard_num: +# Index of shard to use for whole dataset +dataset_shard_idx: + +# 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: 2048 +# Pad inputs so each step uses constant sized buffers +# This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently +pad_to_sequence_len: +# Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true' +sample_packing: +# Set to 'false' if getting errors during eval with sample_packing on. +eval_sample_packing: +# 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: +total_num_tokens: +# Increasing the following values helps with packing, but usually only slightly (<%1.) +# The number of samples packed at a time. +sample_packing_group_size: 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: 200 +sample_pack_sequentially: # Optional[bool]. Whether to pack samples sequentially. + +# whether to concatenate samples during pretraining +pretraining_sample_concatenation: + +curriculum_sampling: # Optional[bool]. Whether to use sequential sampling for curriculum learning + +# Use batch flattening for speedups when not using sample_packing +batch_flattening: + +# Passed through to transformers when loading the model when launched without accelerate +# Use `sequential` when training w/ model parallelism to limit memory +device_map: +# Defines the max memory usage per gpu on the system. Passed through to transformers when loading the model. +max_memory: + +# If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model +adapter: lora +# 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: + +# LoRA hyperparameters +# For more details about the following options, see: +# https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2 +lora_r: 8 +lora_alpha: 16 +lora_dropout: 0.05 +lora_target_modules: + - q_proj + - v_proj +# - k_proj +# - o_proj +# - gate_proj +# - down_proj +# - up_proj +lora_target_linear: # If true, will target all linear modules -# Optional[bool]. Whether to use DoRA. -# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#weight-decomposed-low-rank-adaptation-dora -peft_use_dora: +# List[int] | int. # The layer indices to transform, otherwise, apply to all layers +# https://huggingface.co/docs/peft/v0.15.0/en/package_reference/lora#peft.LoraConfig.layers_to_transform +peft_layers_to_transform: -# Optional[bool]. Whether to use RSLoRA. -# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#rank-stabilized-lora -peft_use_rslora: +# Optional[bool]. Whether to use DoRA. +# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#weight-decomposed-low-rank-adaptation-dora +peft_use_dora: -# Optional[list[tuple[int, int]]]. List of layer indices to replicate. -# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#memory-efficient-layer-replication-with-lora -peft_layer_replication: +# Optional[bool]. Whether to use RSLoRA. +# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#rank-stabilized-lora +peft_use_rslora: -# bool | Literal["gaussian", "eva", "olora", "pissa", "pissa_niter_[number of iters]", "corda", "loftq"] -# How to initialize LoRA weights. Default to True which is MS original implementation. -# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#initialization -peft_init_lora_weights: - -# 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. -# https://github.com/huggingface/peft/issues/334#issuecomment-1561727994 -lora_modules_to_save: -# - embed_tokens -# - lm_head - -lora_fan_in_fan_out: false - -# 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: true -lora_qkv_kernel: true -lora_o_kernel: true - -# LoRA+ hyperparameters -# For more details about the following options, see: -# https://arxiv.org/abs/2402.12354 and `src/axolotl/core/train_builder.py` -loraplus_lr_ratio: # loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4. -loraplus_lr_embedding: # loraplus learning rate for lora embedding layers. Default value is 1e-6. - -peft: - # Configuration options for loftq initialization for LoRA - # https://huggingface.co/docs/peft/developer_guides/quantization#loftq-initialization - loftq_config: - loftq_bits: # typically 4 bits - -# ReLoRA configuration -# Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed -relora_steps: # Number of steps per ReLoRA restart -relora_warmup_steps: # Number of per-restart warmup steps -relora_anneal_steps: # Number of anneal steps for each relora cycle -relora_prune_ratio: # threshold for optimizer magnitude when pruning -relora_cpu_offload: # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings - -# wandb configuration if you're using it -# Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`. -wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb -wandb_project: # Your wandb project name -wandb_entity: # A wandb Team name if using a Team -wandb_watch: -wandb_name: # Set the name of your wandb run -wandb_run_id: # Set the ID of your wandb run -wandb_log_model: # "checkpoint" to log model to wandb Artifacts every `save_steps` or "end" to log only at the end of training - -# mlflow configuration if you're using it -mlflow_tracking_uri: # URI to mlflow -mlflow_experiment_name: # Your experiment name -mlflow_run_name: # Your run name -hf_mlflow_log_artifacts: # set to true to copy each saved checkpoint on each save to mlflow artifact registry - -# Comet configuration if you're using it -# Make sure your `COMET_API_KEY` environment variable is set (recommended) or you login to Comet with `comet login`. -# Check out our documentation for more details https://www.comet.com/docs/v2/api-and-sdk/python-sdk/reference/Experiment-Creation/#comet_ml.start -use_comet: # Enable or disable Comet integration. -comet_api_key: # API key for Comet. Recommended to set via `comet login`. -comet_workspace: # Workspace name in Comet. Defaults to the user's default workspace. -comet_project_name: # Project name in Comet. Defaults to Uncategorized. -comet_experiment_key: # 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_mode: # Create a new experiment ("create") or log to an existing one ("get"). Default ("get_or_create") auto-selects based on configuration. -comet_online: # Set to True to log data to Comet server, or False for offline storage. Default is True. -comet_experiment_config: # Dictionary for additional configuration settings, see the doc for more details. - -# Tensorboard -use_tensorboard: # Optional[bool] - -# Where to save the full-finetuned model to -output_dir: ./completed-model - -# Whether to use torch.compile and which backend to use -# setting to `auto` will enable torch compile when torch>=2.5.1 -torch_compile: # Optional[Union[Literal["auto"], bool]] -torch_compile_backend: # Optional[str] - -# Training hyperparameters - -# If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps. -gradient_accumulation_steps: 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: 2 -eval_batch_size: -num_epochs: 4 -warmup_steps: 100 # cannot use with warmup_ratio -warmup_ratio: 0.05 # cannot use with warmup_steps -learning_rate: 0.00003 -lr_quadratic_warmup: -logging_steps: -eval_steps: # Leave empty to eval at each epoch, integer for every N steps. float for fraction of total steps -evals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps -eval_strategy: # Set to `"no"` to skip evaluation, `"epoch"` at end of each epoch, leave empty to infer from `eval_steps`. -save_strategy: # 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_steps: # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps -saves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps -save_total_limit: # Checkpoints saved at a time -# 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: - -# 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: # Optional[bool] - -# whether to find batch size that fits in memory. Passed to underlying transformers Trainer -auto_find_batch_size: # Optional[bool] - -eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0 -eval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128 -do_causal_lm_eval: # Whether to run causal language model evaluation for metrics in `eval_causal_lm_metrics`. -eval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is ["sacrebleu", "comet", "ter", "chrf", "perplexity"] - -profiler_steps: # 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 +# Optional[list[tuple[int, int]]]. List of layer indices to replicate. +# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#memory-efficient-layer-replication-with-lora +peft_layer_replication: + +# bool | Literal["gaussian", "eva", "olora", "pissa", "pissa_niter_[number of iters]", "corda", "loftq"] +# How to initialize LoRA weights. Default to True which is MS original implementation. +# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#initialization +peft_init_lora_weights: + +# 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. +# https://github.com/huggingface/peft/issues/334#issuecomment-1561727994 +lora_modules_to_save: +# - embed_tokens +# - lm_head + +lora_fan_in_fan_out: false + +# 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: true +lora_qkv_kernel: true +lora_o_kernel: true + +# LoRA+ hyperparameters +# For more details about the following options, see: +# https://arxiv.org/abs/2402.12354 and `src/axolotl/core/train_builder.py` +loraplus_lr_ratio: # loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4. +loraplus_lr_embedding: # loraplus learning rate for lora embedding layers. Default value is 1e-6. + +peft: + # Configuration options for loftq initialization for LoRA + # https://huggingface.co/docs/peft/developer_guides/quantization#loftq-initialization + loftq_config: + loftq_bits: # typically 4 bits + +# ReLoRA configuration +# Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed +relora_steps: # Number of steps per ReLoRA restart +relora_warmup_steps: # Number of per-restart warmup steps +relora_anneal_steps: # Number of anneal steps for each relora cycle +relora_prune_ratio: # threshold for optimizer magnitude when pruning +relora_cpu_offload: # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings + +# wandb configuration if you're using it +# Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`. +wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb +wandb_project: # Your wandb project name +wandb_entity: # A wandb Team name if using a Team +wandb_watch: +wandb_name: # Set the name of your wandb run +wandb_run_id: # Set the ID of your wandb run +wandb_log_model: # "checkpoint" to log model to wandb Artifacts every `save_steps` or "end" to log only at the end of training + +# mlflow configuration if you're using it +mlflow_tracking_uri: # URI to mlflow +mlflow_experiment_name: # Your experiment name +mlflow_run_name: # Your run name +hf_mlflow_log_artifacts: # set to true to copy each saved checkpoint on each save to mlflow artifact registry + +# Comet configuration if you're using it +# Make sure your `COMET_API_KEY` environment variable is set (recommended) or you login to Comet with `comet login`. +# Check out our documentation for more details https://www.comet.com/docs/v2/api-and-sdk/python-sdk/reference/Experiment-Creation/#comet_ml.start +use_comet: # Enable or disable Comet integration. +comet_api_key: # API key for Comet. Recommended to set via `comet login`. +comet_workspace: # Workspace name in Comet. Defaults to the user's default workspace. +comet_project_name: # Project name in Comet. Defaults to Uncategorized. +comet_experiment_key: # 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_mode: # Create a new experiment ("create") or log to an existing one ("get"). Default ("get_or_create") auto-selects based on configuration. +comet_online: # Set to True to log data to Comet server, or False for offline storage. Default is True. +comet_experiment_config: # Dictionary for additional configuration settings, see the doc for more details. + +# Tensorboard +use_tensorboard: # Optional[bool] + +# Where to save the full-finetuned model to +output_dir: ./completed-model + +# Whether to use torch.compile and which backend to use +# setting to `auto` will enable torch compile when torch>=2.5.1 +torch_compile: # Optional[Union[Literal["auto"], bool]] +torch_compile_backend: # Optional[str] + +# Training hyperparameters + +# If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps. +gradient_accumulation_steps: 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: 2 +eval_batch_size: +num_epochs: 4 +warmup_steps: 100 # cannot use with warmup_ratio +warmup_ratio: 0.05 # cannot use with warmup_steps +learning_rate: 0.00003 +lr_quadratic_warmup: +logging_steps: +eval_steps: # Leave empty to eval at each epoch, integer for every N steps. float for fraction of total steps +evals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps +eval_strategy: # Set to `"no"` to skip evaluation, `"epoch"` at end of each epoch, leave empty to infer from `eval_steps`. +save_strategy: # 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_steps: # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps +saves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps +save_total_limit: # Checkpoints saved at a time +# 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: + +# 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: # Optional[bool] + +# whether to find batch size that fits in memory. Passed to underlying transformers Trainer +auto_find_batch_size: # Optional[bool] + +eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0 +eval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128 +do_causal_lm_eval: # Whether to run causal language model evaluation for metrics in `eval_causal_lm_metrics`. +eval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is ["sacrebleu", "comet", "ter", "chrf", "perplexity"] -loss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training) -loss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3) - -# Save model as safetensors (require safetensors package) -save_safetensors: - -# Whether to mask out or include the human's prompt from the training labels -train_on_inputs: 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: false - -# Whether to use gradient checkpointing. Available options are: true, false, "offload". -# https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing -gradient_checkpointing: false -# additional kwargs to pass to the trainer for gradient checkpointing -# gradient_checkpointing_kwargs: -# use_reentrant: true - -# Stop training after this many evaluation losses have increased in a row -# https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback -early_stopping_patience: 3 +profiler_steps: # 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 + +loss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training) +loss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3) + +# Save model as safetensors (require safetensors package) +save_safetensors: + +# Whether to mask out or include the human's prompt from the training labels +train_on_inputs: 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: false + +# Whether to use gradient checkpointing. Available options are: true, false, "offload". +# https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing +gradient_checkpointing: false +# additional kwargs to pass to the trainer for gradient checkpointing +# gradient_checkpointing_kwargs: +# use_reentrant: true -# Specify a scheduler and kwargs to use with the optimizer -lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | empty for cosine -lr_scheduler_kwargs: -cosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr -cosine_constant_lr_ratio: # 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 (https://arxiv.org/pdf/2308.04014.pdf) - -# For one_cycle optim -lr_div_factor: # Learning rate div factor - -# Specify optimizer -# Valid values are driven by the Transformers OptimizerNames class, see: -# https://github.com/huggingface/transformers/blob/cbf924b76c03828101a34069a96d209314114fd5/src/transformers/training_args.py#L144-L189 -# -# Note that not all optimizers may be available in your environment, ex: 'adamw_anyprecision' is part of -# torchdistx, 'adamw_bnb_8bit' is part of bnb.optim.Adam8bit, etc. When in doubt, it is recommended to start with the optimizer used -# in the examples/ for your model and fine-tuning use case. +# Stop training after this many evaluation losses have increased in a row +# https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback +early_stopping_patience: 3 + +# Specify a scheduler and kwargs to use with the optimizer +lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | empty for cosine +lr_scheduler_kwargs: +cosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr +cosine_constant_lr_ratio: # 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 (https://arxiv.org/pdf/2308.04014.pdf) + +# For one_cycle optim +lr_div_factor: # Learning rate div factor + +# Specify optimizer +# Valid values are driven by the Transformers OptimizerNames class, see: +# https://github.com/huggingface/transformers/blob/cbf924b76c03828101a34069a96d209314114fd5/src/transformers/training_args.py#L144-L189 # -# Valid values for 'optimizer' include: -# - adamw_torch -# - adamw_torch_fused -# - adamw_torch_xla -# - adamw_torch_npu_fused -# - adamw_apex_fused -# - adopt_adamw (an EXPERIMENTAL optimizer, only for torch version >= 2.5.1) -# - adafactor -# - adamw_anyprecision -# - adamw_torch_4bit -# - ademamix -# - sgd -# - adagrad -# - adamw_bnb_8bit -# - adamw_8bit # alias for adamw_bnb_8bit -# - ademamix_8bit -# - lion_8bit -# - lion_32bit -# - paged_adamw_32bit -# - paged_adamw_8bit -# - paged_ademamix_32bit -# - paged_ademamix_8bit -# - paged_lion_32bit -# - paged_lion_8bit -# - rmsprop -# - rmsprop_bnb -# - rmsprop_bnb_8bit -# - rmsprop_bnb_32bit -# - galore_adamw -# - galore_adamw_8bit -# - galore_adafactor -# - galore_adamw_layerwise -# - galore_adamw_8bit_layerwise -# - galore_adafactor_layerwise -# - lomo -# - adalomo -# - grokadamw -# - schedule_free_adamw -# - schedule_free_sgd -# - apollo_adamw -# - apollo_adamw_layerwise -# -# Additional custom optimizers include: -# - optimi_adamw -# - ao_adamw_8bit -# - ao_adamw_fp8 -optimizer: -# Dictionary of arguments to pass to the optimizer -optim_args: -# For Galore Optimizers the following optim_args are available -# rank: # type: int -# update_proj_gap # type: int -# scale # type: float -# proj_type: # type: str, default = std - -# 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: -# - self_attn # for llama -# - mlp - -# Specify weight decay -weight_decay: -# adamw hyperparams -adam_beta1: -adam_beta2: -adam_epsilon: -# Gradient clipping max norm -max_grad_norm: - -# Augmentation techniques -# 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: - -# Optional[bool]. Whether to bettertransformers -flash_optimum: - -# Note: Only one of the following attention patches can be used at a time. -# For example, if you set `xformers_attention` to `true`, do not set `flash_attention` to `true`. - -# Optional[bool]. Whether to use xformers attention patch https://github.com/facebookresearch/xformers: -xformers_attention: -# Optional[bool]. Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention: -flash_attention: -flash_attn_cross_entropy: # Optional[bool]. Whether to use flash-attention cross entropy implementation - advanced use only -flash_attn_rms_norm: # Optional[bool]. Whether to use flash-attention rms norm implementation - advanced use only -flash_attn_fuse_qkv: # Optional[bool]. Whether to fuse QKV into a single operation -flash_attn_fuse_mlp: # Optional[bool]. Whether to fuse part of the MLP into a single operation -# Optional[bool]. Whether to use scaled-dot-product attention -# https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html -sdp_attention: -# Optional[bool]. Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf -s2_attention: - -# Optional[bool]. Whether to use low_cpu_mem_usage -low_cpu_mem_usage: -# Optional[str]. Resume from a specific checkpoint dir -resume_from_checkpoint: -# Optional[bool]. If resume_from_checkpoint isn't set and you simply want it to start where it left off. -# Be careful with this being turned on between different models. -auto_resume_from_checkpoints: false - -## Multimodal section -# int | tuple[int, int] | None . Size to resize images to, width x height. -# Will read from model/processor config if not set. -image_size: -# str. Algorithm to use for image resizing. "bilinear", "bicubic", "lanczos". Default is "bilinear". -image_resize_algorithm: 'bilinear' -## End of multimodal section - -# Don't mess with this, it's here for accelerate and torchrun -local_rank: - -# Add or change special tokens. -# If you add tokens here, you don't need to add them to the `tokens` list. -special_tokens: - # bos_token: "<s>" - # eos_token: "</s>" - # unk_token: "<unk>" - # pad_token: "[PAD]" - -# Optional[list[str]]. Add extra tokens to the tokenizer. -tokens: - # - "<|startoftext|>" - # - "<|endoftext|>" - -# 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] -# 128041: "<|im_start|>" -# 128042: "<|im_end|>" - -# FSDP -fsdp: -fsdp_config: +# Note that not all optimizers may be available in your environment, ex: 'adamw_anyprecision' is part of +# torchdistx, 'adamw_bnb_8bit' is part of bnb.optim.Adam8bit, etc. When in doubt, it is recommended to start with the optimizer used +# in the examples/ for your model and fine-tuning use case. +# +# Valid values for 'optimizer' include: +# - adamw_torch +# - adamw_torch_fused +# - adamw_torch_xla +# - adamw_torch_npu_fused +# - adamw_apex_fused +# - adopt_adamw (an EXPERIMENTAL optimizer, only for torch version >= 2.5.1) +# - adafactor +# - adamw_anyprecision +# - adamw_torch_4bit +# - ademamix +# - sgd +# - adagrad +# - adamw_bnb_8bit +# - adamw_8bit # alias for adamw_bnb_8bit +# - ademamix_8bit +# - lion_8bit +# - lion_32bit +# - paged_adamw_32bit +# - paged_adamw_8bit +# - paged_ademamix_32bit +# - paged_ademamix_8bit +# - paged_lion_32bit +# - paged_lion_8bit +# - rmsprop +# - rmsprop_bnb +# - rmsprop_bnb_8bit +# - rmsprop_bnb_32bit +# - galore_adamw +# - galore_adamw_8bit +# - galore_adafactor +# - galore_adamw_layerwise +# - galore_adamw_8bit_layerwise +# - galore_adafactor_layerwise +# - lomo +# - adalomo +# - grokadamw +# - schedule_free_adamw +# - schedule_free_sgd +# - apollo_adamw +# - apollo_adamw_layerwise +# +# Additional custom optimizers include: +# - optimi_adamw +# - ao_adamw_8bit +# - ao_adamw_fp8 +optimizer: +# Dictionary of arguments to pass to the optimizer +optim_args: +# For Galore Optimizers the following optim_args are available +# rank: # type: int +# update_proj_gap # type: int +# scale # type: float +# proj_type: # type: str, default = std + +# 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: +# - self_attn # for llama +# - mlp + +# Specify weight decay +weight_decay: +# adamw hyperparams +adam_beta1: +adam_beta2: +adam_epsilon: +# Gradient clipping max norm +max_grad_norm: + +# Augmentation techniques +# 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: + +# Optional[bool]. Whether to bettertransformers +flash_optimum: + +# Note: Only one of the following attention patches can be used at a time. +# For example, if you set `xformers_attention` to `true`, do not set `flash_attention` to `true`. + +# Optional[bool]. Whether to use xformers attention patch https://github.com/facebookresearch/xformers: +xformers_attention: +# Optional[bool]. Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention: +flash_attention: +flash_attn_cross_entropy: # Optional[bool]. Whether to use flash-attention cross entropy implementation - advanced use only +flash_attn_rms_norm: # Optional[bool]. Whether to use flash-attention rms norm implementation - advanced use only +flash_attn_fuse_qkv: # Optional[bool]. Whether to fuse QKV into a single operation +flash_attn_fuse_mlp: # Optional[bool]. Whether to fuse part of the MLP into a single operation +# Optional[bool]. Whether to use scaled-dot-product attention +# https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html +sdp_attention: +# Optional[bool]. Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf +s2_attention: + +# Optional[bool]. Whether to use low_cpu_mem_usage +low_cpu_mem_usage: +# Optional[str]. Resume from a specific checkpoint dir +resume_from_checkpoint: +# Optional[bool]. If resume_from_checkpoint isn't set and you simply want it to start where it left off. +# Be careful with this being turned on between different models. +auto_resume_from_checkpoints: false + +## Multimodal section +# int | tuple[int, int] | None . Size to resize images to, width x height. +# Will read from model/processor config if not set. +image_size: +# str. Algorithm to use for image resizing. "bilinear", "bicubic", "lanczos". Default is "bilinear". +image_resize_algorithm: 'bilinear' +## End of multimodal section + +# Don't mess with this, it's here for accelerate and torchrun +local_rank: + +# Add or change special tokens. +# If you add tokens here, you don't need to add them to the `tokens` list. +special_tokens: + # bos_token: "<s>" + # eos_token: "</s>" + # unk_token: "<unk>" + # pad_token: "[PAD]" + +# Optional[list[str]]. Add extra tokens to the tokenizer. +tokens: + # - "<|startoftext|>" + # - "<|endoftext|>" + +# 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] +# 128041: "<|im_start|>" +# 128042: "<|im_end|>" -# Deepspeed config path. e.g., deepspeed_configs/zero3.json -deepspeed: - -# Advanced DDP Arguments -ddp_timeout: -ddp_bucket_cap_mb: -ddp_broadcast_buffers: - -# Sequence parallelism -# 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. -sequence_parallel_degree: -# 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: 1 -# 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: - -# Path to torch distx for optim 'adamw_anyprecision' -torchdistx_path: - -# Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize -pretraining_dataset: - -# Debug mode -debug: - -# Seed -seed: - -# Allow overwrite yml config using from cli -strict: +# FSDP +fsdp: +fsdp_config: + +# Deepspeed config path. e.g., deepspeed_configs/zero3.json +deepspeed: + +# Advanced DDP Arguments +ddp_timeout: +ddp_bucket_cap_mb: +ddp_broadcast_buffers: + +# Sequence parallelism +# 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. +sequence_parallel_degree: +# 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: 1 +# 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: + +# Path to torch distx for optim 'adamw_anyprecision' +torchdistx_path: + +# Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize +pretraining_dataset: + +# Debug mode +debug: + +# Seed +seed: + +# Allow overwrite yml config using from cli +strict: diff --git a/search.json b/search.json index c64e3c009..3cb1dca79 100644 --- a/search.json +++ b/search.json @@ -1202,7 +1202,7 @@ "href": "docs/config.html", "title": "Config Reference", "section": "", - "text": "# This is the huggingface model that contains *.pt, *.safetensors, or *.bin files\n# This can also be a relative path to a model on disk\nbase_model: ./llama-7b-hf\n# You can specify an ignore pattern if the model repo contains more than 1 model type (*.pt, etc)\nbase_model_ignore_patterns:\n# If the base_model repo on hf hub doesn't include configuration .json files,\n# You can set that here, or leave this empty to default to base_model\nbase_model_config: ./llama-7b-hf\n# You can specify to choose a specific model revision from huggingface hub\nrevision_of_model:\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:\n# If you want to specify the type of model to load, AutoModelForCausalLM is a good choice too\nmodel_type: AutoModelForCausalLM\n# Corresponding tokenizer for the model AutoTokenizer is a good choice\ntokenizer_type: AutoTokenizer\n# Trust remote code for untrusted source\ntrust_remote_code:\n# use_fast option for tokenizer loading from_pretrained, default to True\ntokenizer_use_fast:\n# Whether to use the legacy tokenizer setting, defaults to True\ntokenizer_legacy:\n# Resize the model embeddings when new tokens are added to multiples of 32\n# This is reported to improve training speed on some models\nresize_token_embeddings_to_32x:\n# Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.\nshrink_embeddings:\n# Whether to load the model with randomly initialized weights. Useful for\n# pre-training a model from scratch or debugging purposes.\nrandom_init_weights:\n\n# (Internal use only)\n# Used to identify which the model is based on\nis_falcon_derived_model:\nis_llama_derived_model:\nis_qwen_derived_model:\n# Please note that if you set this to true, `padding_side` will be set to \"left\" by default\nis_mistral_derived_model:\n\n# optional overrides to the base model configuration\noverrides_of_model_config:\n # RoPE Scaling https://github.com/huggingface/transformers/pull/24653\n rope_scaling:\n type: # linear | dynamic\n factor: # float\n\n# optional overrides the base model loading from_pretrained\noverrides_of_model_kwargs:\n # use_cache: False\n\n# optional overrides to the bnb 4bit quantization configuration\n# https://huggingface.co/docs/transformers/main/main_classes/quantization#transformers.BitsAndBytesConfig\nbnb_config_kwargs:\n # These are default values\n llm_int8_has_fp16_weight: false\n bnb_4bit_quant_type: nf4\n bnb_4bit_use_double_quant: true\n\n\n# Whether you are training a 4-bit GPTQ quantized model\ngptq: true\n\n# This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer\nload_in_8bit: true\n# Use bitsandbytes 4 bit\nload_in_4bit:\n\n# Use CUDA bf16\nbf16: true # bool or 'full' for `bf16_full_eval`. require >=ampere\n# Use CUDA fp16\nfp16: true\n# Use CUDA tf32\ntf32: true # require >=ampere\n\n# No AMP (automatic mixed precision)\nbfloat16: true # require >=ampere\nfloat16: true\n\n# Limit the memory for all available GPUs to this amount (if an integer, expressed in gigabytes); default: unset\ngpu_memory_limit: 20GiB\n# 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\nlora_on_cpu: true\n\n# List[str]. Add plugins to extend the pipeline.\n# See `src/axolotl/integrations` for the available plugins or doc below for more details.\n# https://docs.axolotl.ai/docs/custom_integrations.html\nplugins:\n # - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin\n\n# A list of one or more datasets to finetune the model with\ndatasets:\n # HuggingFace dataset repo | s3://,gs:// path | \"json\" for local dataset, make sure to fill data_files\n - path: vicgalle/alpaca-gpt4\n # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]\n type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>\n ds_type: # Optional[str] (json|arrow|parquet|text|csv) defines the datatype when path is a file\n data_files: # Optional[str] path to source data files\n\n shards: # Optional[int] split dataset into N pieces (use with shards_idx)\n shards_idx: # Optional[int] = 0 the index of sharded dataset to use\n\n preprocess_shards: # Optional[int] process dataset in N sequential chunks for memory efficiency (exclusive with `shards`)\n\n name: # Optional[str] name of dataset configuration to load\n split: train # Optional[str] name of dataset split to load from\n revision: # Optional[str] 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.\n trust_remote_code: # Optional[bool] Trust remote code for untrusted source\n\n # Custom user instruction prompt\n - path: repo\n type:\n # The below are defaults. only set what's needed if you use a different column name.\n system_prompt: \"\"\n system_format: \"{system}\"\n field_system: system\n field_instruction: instruction\n field_input: input\n field_output: output\n\n # Customizable to be single line or multi-line\n # Use {instruction}/{input} as key to be replaced\n # 'format' can include {input}\n format: |-\n User: {instruction} {input}\n Assistant:\n # 'no_input_format' cannot include {input}\n no_input_format: \"{instruction} \"\n\n # For `completion` datsets only, uses the provided field instead of `text` column\n field:\n\n # Using chat template\n - path: ...\n # Set type to `chat_template` to use this strategy\n type: chat_template\n # Specify the name of the chat template to use\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 tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default.\n # - 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\n # - 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.\n # - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.\n chat_template: tokenizer_default\n\n # Custom jinja chat template. Used only if `chat_template: jinja` or empty.\n chat_template_jinja:\n\n # Key containing the messages (default: \"messages\")\n field_messages: messages\n\n # Key containing the system message (default: \"system\")\n # If the system message is not present in the dataset sample, it will be loaded from the field_system property.\n field_system: system\n\n # Mapping of properties from the input dataset to the chat template.\n # (default: message_property_mappings={'role':'role', 'content':'content'})\n # If a property exists in the template but not in this mapping, the system will attempt\n # to load it directly from the message using the property name as the key.\n # Example: In the mapping below, 'from' is loaded from input dataset and used as 'role',\n # while 'value' is loaded and used as 'content' in the chat template.\n message_property_mappings:\n role: from\n content: value\n # ...\n\n # Optional[Dict[str, List]]. Roles mapping in the messages.\n # The format is {target_role: [source_roles]}. All source roles will be mapped to the target role.\n # The default is:\n roles:\n user: [\"human\", \"user\"]\n assistant: [\"gpt\", \"assistant\"]\n system: [\"system\"]\n tool: [\"tool\"]\n\n # Optional[bool]. 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 you wish to,\n # we recommend using a custom jinja template with the default system message removed or\n # adding a system turn with empty content.\n drop_system_message:\n\n # IMPORTANT: The following fields determine which parts of the conversation to train on.\n # Priority order: message_field_training > message_field_training_detail > train_on_inputs or role in roles_to_train\n # See examples at `docs/dataset-formats/conversation.qmd`\n # Note: If the below 5 fields are empty, defaults to training only on the last message.\n\n # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.\n roles_to_train: [\"assistant\"] # default\n # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are:\n # - all: train on all EOS tokens\n # - turn (default): train on the EOS token at the end of each trainable turn\n # - last: train on the last EOS token in the conversation\n # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.\n train_on_eos: turn\n # Optional[str]. Which EOT (End-of-Turn) tokens to train on in the conversation. Possible values are:\n # - all: train on all EOT tokens\n # - turn: train on the EOT token at the end of each trainable turn\n # - last: train on the last EOT token in the conversation\n # If not specified, defaults to the value of train_on_eos for backward compatibility.\n train_on_eot:\n # 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`.\n message_field_training: training\n # The key in the message turn that contains the training details. Useful to selectively train on certain tokens in a turn.\n # 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).\n message_field_training_detail: train_detail\n\n\n# If false, the datasets will not be shuffled and will keep their original order in `datasets`.\n# The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.\nshuffle_merged_datasets: true\n\nDeduplicates datasets and test_datasets with identical entries.\ndataset_exact_deduplication: true\n\n# A list of one or more datasets to eval the model with.\n# You can use either test_datasets, or val_set_size, but not both.\ntest_datasets:\n - path: /workspace/data/eval.jsonl\n ds_type: json\n # You need to specify a split. For \"json\" datasets the default split is called \"train\".\n split: train\n type: completion\n data_files:\n - /workspace/data/eval.jsonl\n\n# use RL training: 'dpo', 'ipo', 'kto', 'simpo', 'orpo', 'grpo'\nrl:\nrl_beta: # Optional[float]. The beta parameter for the RL training.\n\n# dpo\ndpo_use_weighting: # Optional[bool]. Whether to perform weighting.\nrpo_alpha: # Optional[float]. Weighting of NLL term in loss from RPO paper.\n\n# orpo\norpo_alpha: 0.1 # Parameter controlling the relative ratio loss weight in the ORPO loss. Passed to `beta` in `ORPOConfig` due to trl mapping.\n\n# kto\nkto_desirable_weight: # Optional[float]. Factor for desirable loss term in KTO loss.\nkto_undesirable_weight: # Optional[float]. Factor for undesirable loss term in KTO loss.\n\n# simpo\ncpo_alpha: 1.0 # Weight of the BC regularizer\nsimpo_gamma: 0.5 # Target reward margin for the SimPO loss\n\n# grpo\ntrl:\n use_vllm: # Optional[bool]. Whether to use VLLM for RL training.\n vllm_server_host: # Optional[str]. Host of the vLLM server to connect to.\n vllm_server_port: # Optional[int]. Port of the vLLM server to connect to.\n vllm_server_timeout: # Optional[int]. Total timeout (in seconds) to wait for the vLLM server to respond.\n vllm_guided_decoding_regex: # Optional[str]. Regex for vLLM guided decoding.\n\n beta: # Optional[float]. Beta parameter for the RL training. Same as `rl_beta`. Use\n max_completion_length: # Optional[int]. Maximum length of the completion for RL training.\n\n reward_funcs: # Optional[list[str]]. List of reward functions to load. Paths must be importable from current dir.\n reward_weights: # Optional[list[float]]. List of reward weights for the reward functions.\n\n num_generations: # Optional[int]. Number of generations to sample.\n log_completions: # Optional[bool]. Whether to log completions.\n\n sync_ref_model: # Optional[bool]. Whether to sync the reference model.\n ref_model_mixup_alpha: # Optional[float]. Mixup alpha for the reference model.\n ref_model_sync_steps: # Optional[int]. Sync steps for the reference model.\n\n\n# reward modelling: `True` or `False`\nreward_model:\n\n# process reward modelling: `True` or `False`\nprocess_reward_model:\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 tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default value.\n# - 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\n# - 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.\n# - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.\n# The selected chat template will be saved to the tokenizer_config.json for easier inferencing\n# Note: It is recommended to set train_on_inputs to true when using a chat template that is different from the model's default chat template.\nchat_template: tokenizer_default\n# custom jinja template 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.\nchat_template_jinja: null\n# Optional[List[str]]. Custom EOT (End-of-Turn) tokens to mask/unmask during training.\n# These tokens mark the boundaries between conversation turns.\n# For example: [\"/INST\", \"</s>\", \"[/SYSTEM_PROMPT]\"]\n# If not specified, defaults to just the model's eos_token.\n# This is useful for templates that use multiple delimiter tokens.\neot_tokens:\n # - \"</s>\"\n # - \"[/INST]\"\n # - \"[/SYSTEM_PROMPT]\"\n# Changes the default system message\ndefault_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml.\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: data/last_run_prepared\n# Push prepared dataset to hub\npush_dataset_to_hub: # Optional[str] repo_org/repo_name\n# The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`\n# if not set.\ndataset_processes: # defaults to os.cpu_count() if not set\n# Keep dataset in memory while preprocessing\n# Only needed if cached dataset is taking too much storage\ndataset_keep_in_memory:\n# push checkpoints to hub\nhub_model_id: # private repo path to push finetuned model\n# how to push checkpoints to hub\n# https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy\nhub_strategy:\n# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets\n# Required to be true when used in combination with `push_dataset_to_hub`\nhf_use_auth_token: # boolean\n# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval.\nval_set_size: 0.04\n# Num shards for whole dataset\ndataset_shard_num:\n# Index of shard to use for whole dataset\ndataset_shard_idx:\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: 2048\n# Pad inputs so each step uses constant sized buffers\n# This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently\npad_to_sequence_len:\n# Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'\nsample_packing:\n# Set to 'false' if getting errors during eval with sample_packing on.\neval_sample_packing:\n# You can set these packing optimizations AFTER starting a training at least once.\n# The trainer will provide recommended values for these values.\nsample_packing_eff_est:\ntotal_num_tokens:\n# Increasing the following values helps with packing, but usually only slightly (<%1.)\n# The number of samples packed at a time.\nsample_packing_group_size: 100000\n# The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.\nsample_packing_bin_size: 200\nsample_pack_sequentially: # Optional[bool]. Whether to pack samples sequentially.\n\n# whether to concatenate samples during pretraining\npretraining_sample_concatenation:\n\ncurriculum_sampling: # Optional[bool]. Whether to use sequential sampling for curriculum learning\n\n# Use batch flattening for speedups when not using sample_packing\nbatch_flattening:\n\n# Passed through to transformers when loading the model when launched without accelerate\n# Use `sequential` when training w/ model parallelism to limit memory\ndevice_map:\n# Defines the max memory usage per gpu on the system. Passed through to transformers when loading the model.\nmax_memory:\n\n# If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model\nadapter: lora\n# If you already have a lora model trained that you want to load, put that here.\n# This means after training, if you want to test the model, you should set this to the value of `output_dir`.\n# Note that if you merge an adapter to the base model, a new subdirectory `merged` will be created under the `output_dir`.\nlora_model_dir:\n\n# LoRA hyperparameters\n# For more details about the following options, see:\n# https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2\nlora_r: 8\nlora_alpha: 16\nlora_dropout: 0.05\nlora_target_modules:\n - q_proj\n - v_proj\n# - k_proj\n# - o_proj\n# - gate_proj\n# - down_proj\n# - up_proj\nlora_target_linear: # If true, will target all linear modules\n\n# List[int] | int. # The layer indices to transform, otherwise, apply to all layers\n# https://huggingface.co/docs/peft/v0.15.0/en/package_reference/lora#peft.LoraConfig.layers_to_transform\npeft_layers_to_transform:\n\n# Optional[bool]. Whether to use DoRA.\n# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#weight-decomposed-low-rank-adaptation-dora\npeft_use_dora:\n\n# Optional[bool]. Whether to use RSLoRA.\n# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#rank-stabilized-lora\npeft_use_rslora:\n\n# Optional[list[tuple[int, int]]]. List of layer indices to replicate.\n# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#memory-efficient-layer-replication-with-lora\npeft_layer_replication:\n\n# bool | Literal[\"gaussian\", \"eva\", \"olora\", \"pissa\", \"pissa_niter_[number of iters]\", \"corda\", \"loftq\"]\n# How to initialize LoRA weights. Default to True which is MS original implementation.\n# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#initialization\npeft_init_lora_weights:\n\n# If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.\n# For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.\n# `embed_tokens` converts tokens to embeddings, and `lm_head` converts embeddings to token probabilities.\n# https://github.com/huggingface/peft/issues/334#issuecomment-1561727994\nlora_modules_to_save:\n# - embed_tokens\n# - lm_head\n\nlora_fan_in_fan_out: false\n\n# Apply custom LoRA autograd functions and activation function Triton kernels for\n# speed and memory savings\n# See: https://docs.axolotl.ai/docs/lora_optims.html\nlora_mlp_kernel: true\nlora_qkv_kernel: true\nlora_o_kernel: true\n\n# LoRA+ hyperparameters\n# For more details about the following options, see:\n# https://arxiv.org/abs/2402.12354 and `src/axolotl/core/train_builder.py`\nloraplus_lr_ratio: # loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4.\nloraplus_lr_embedding: # loraplus learning rate for lora embedding layers. Default value is 1e-6.\n\npeft:\n # Configuration options for loftq initialization for LoRA\n # https://huggingface.co/docs/peft/developer_guides/quantization#loftq-initialization\n loftq_config:\n loftq_bits: # typically 4 bits\n\n# ReLoRA configuration\n# Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed\nrelora_steps: # Number of steps per ReLoRA restart\nrelora_warmup_steps: # Number of per-restart warmup steps\nrelora_anneal_steps: # Number of anneal steps for each relora cycle\nrelora_prune_ratio: # threshold for optimizer magnitude when pruning\nrelora_cpu_offload: # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings\n\n# wandb configuration if you're using it\n# Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.\nwandb_mode: # \"offline\" to save run metadata locally and not sync to the server, \"disabled\" to turn off wandb\nwandb_project: # Your wandb project name\nwandb_entity: # A wandb Team name if using a Team\nwandb_watch:\nwandb_name: # Set the name of your wandb run\nwandb_run_id: # Set the ID of your wandb run\nwandb_log_model: # \"checkpoint\" to log model to wandb Artifacts every `save_steps` or \"end\" to log only at the end of training\n\n# mlflow configuration if you're using it\nmlflow_tracking_uri: # URI to mlflow\nmlflow_experiment_name: # Your experiment name\nmlflow_run_name: # Your run name\nhf_mlflow_log_artifacts: # set to true to copy each saved checkpoint on each save to mlflow artifact registry\n\n# Comet configuration if you're using it\n# Make sure your `COMET_API_KEY` environment variable is set (recommended) or you login to Comet with `comet login`.\n# Check out our documentation for more details https://www.comet.com/docs/v2/api-and-sdk/python-sdk/reference/Experiment-Creation/#comet_ml.start\nuse_comet: # Enable or disable Comet integration.\ncomet_api_key: # API key for Comet. Recommended to set via `comet login`.\ncomet_workspace: # Workspace name in Comet. Defaults to the user's default workspace.\ncomet_project_name: # Project name in Comet. Defaults to Uncategorized.\ncomet_experiment_key: # Identifier for the experiment. Used to append data to an existing experiment or control the key of new experiments. Default to a random key.\ncomet_mode: # Create a new experiment (\"create\") or log to an existing one (\"get\"). Default (\"get_or_create\") auto-selects based on configuration.\ncomet_online: # Set to True to log data to Comet server, or False for offline storage. Default is True.\ncomet_experiment_config: # Dictionary for additional configuration settings, see the doc for more details.\n\n# Tensorboard\nuse_tensorboard: # Optional[bool]\n\n# Where to save the full-finetuned model to\noutput_dir: ./completed-model\n\n# Whether to use torch.compile and which backend to use\n# setting to `auto` will enable torch compile when torch>=2.5.1\ntorch_compile: # Optional[Union[Literal[\"auto\"], bool]]\ntorch_compile_backend: # Optional[str]\n\n# Training hyperparameters\n\n# If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps.\ngradient_accumulation_steps: 1\n# The number of samples to include in each batch. This is the number of samples sent to each GPU.\n# Batch size per gpu = micro_batch_size * gradient_accumulation_steps\nmicro_batch_size: 2\neval_batch_size:\nnum_epochs: 4\nwarmup_steps: 100 # cannot use with warmup_ratio\nwarmup_ratio: 0.05 # cannot use with warmup_steps\nlearning_rate: 0.00003\nlr_quadratic_warmup:\nlogging_steps:\neval_steps: # Leave empty to eval at each epoch, integer for every N steps. float for fraction of total steps\nevals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps\neval_strategy: # Set to `\"no\"` to skip evaluation, `\"epoch\"` at end of each epoch, leave empty to infer from `eval_steps`.\nsave_strategy: # 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`.\nsave_steps: # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps\nsaves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps\nsave_total_limit: # Checkpoints saved at a time\n# Maximum number of iterations to train for. It precedes num_epochs which means that\n# if both are set, num_epochs will not be guaranteed.\n# e.g., when 1 epoch is 1000 steps => `num_epochs: 2` and `max_steps: 100` will train for 100 steps\nmax_steps:\n\n# 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.\ninclude_tokens_per_second: # Optional[bool]\n\n# whether to find batch size that fits in memory. Passed to underlying transformers Trainer\nauto_find_batch_size: # Optional[bool]\n\neval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0\neval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128\ndo_causal_lm_eval: # Whether to run causal language model evaluation for metrics in `eval_causal_lm_metrics`.\neval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is [\"sacrebleu\", \"comet\", \"ter\", \"chrf\", \"perplexity\"]\n\nprofiler_steps: # enable the pytorch profiler to capture the first N steps of training to the output_dir.\n # see https://pytorch.org/blog/understanding-gpu-memory-1/ for more information\n # snapshots can be visualized @ https://pytorch.org/memory_viz\n\nloss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)\nloss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3)\n\n# Save model as safetensors (require safetensors package)\nsave_safetensors:\n\n# Whether to mask out or include the human's prompt from the training labels\ntrain_on_inputs: false\n# Group similarly sized data to minimize padding.\n# May be slower to start, as it must download and sort the entire dataset.\n# Note that training loss may have an oscillating pattern with this enabled.\ngroup_by_length: false\n\n# Whether to use gradient checkpointing. Available options are: true, false, \"offload\".\n# https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing\ngradient_checkpointing: false\n# additional kwargs to pass to the trainer for gradient checkpointing\n# gradient_checkpointing_kwargs:\n# use_reentrant: true\n\n# Stop training after this many evaluation losses have increased in a row\n# https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback\nearly_stopping_patience: 3\n\n# Specify a scheduler and kwargs to use with the optimizer\nlr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | empty for cosine\nlr_scheduler_kwargs:\ncosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr\ncosine_constant_lr_ratio: # 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 (https://arxiv.org/pdf/2308.04014.pdf)\n\n# For one_cycle optim\nlr_div_factor: # Learning rate div factor\n\n# Specify optimizer\n# Valid values are driven by the Transformers OptimizerNames class, see:\n# https://github.com/huggingface/transformers/blob/cbf924b76c03828101a34069a96d209314114fd5/src/transformers/training_args.py#L144-L189\n#\n# Note that not all optimizers may be available in your environment, ex: 'adamw_anyprecision' is part of\n# torchdistx, 'adamw_bnb_8bit' is part of bnb.optim.Adam8bit, etc. When in doubt, it is recommended to start with the optimizer used\n# in the examples/ for your model and fine-tuning use case.\n#\n# Valid values for 'optimizer' include:\n# - adamw_torch\n# - adamw_torch_fused\n# - adamw_torch_xla\n# - adamw_torch_npu_fused\n# - adamw_apex_fused\n# - adopt_adamw (an EXPERIMENTAL optimizer, only for torch version >= 2.5.1)\n# - adafactor\n# - adamw_anyprecision\n# - adamw_torch_4bit\n# - ademamix\n# - sgd\n# - adagrad\n# - adamw_bnb_8bit\n# - adamw_8bit # alias for adamw_bnb_8bit\n# - ademamix_8bit\n# - lion_8bit\n# - lion_32bit\n# - paged_adamw_32bit\n# - paged_adamw_8bit\n# - paged_ademamix_32bit\n# - paged_ademamix_8bit\n# - paged_lion_32bit\n# - paged_lion_8bit\n# - rmsprop\n# - rmsprop_bnb\n# - rmsprop_bnb_8bit\n# - rmsprop_bnb_32bit\n# - galore_adamw\n# - galore_adamw_8bit\n# - galore_adafactor\n# - galore_adamw_layerwise\n# - galore_adamw_8bit_layerwise\n# - galore_adafactor_layerwise\n# - lomo\n# - adalomo\n# - grokadamw\n# - schedule_free_adamw\n# - schedule_free_sgd\n# - apollo_adamw\n# - apollo_adamw_layerwise\n#\n# Additional custom optimizers include:\n# - optimi_adamw\n# - ao_adamw_8bit\n# - ao_adamw_fp8\noptimizer:\n# Dictionary of arguments to pass to the optimizer\noptim_args:\n# For Galore Optimizers the following optim_args are available\n# rank: # type: int\n# update_proj_gap # type: int\n# scale # type: float\n# proj_type: # type: str, default = std\n\n# 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\noptim_target_modules:\n# - self_attn # for llama\n# - mlp\n\n# Specify weight decay\nweight_decay:\n# adamw hyperparams\nadam_beta1:\nadam_beta2:\nadam_epsilon:\n# Gradient clipping max norm\nmax_grad_norm:\n\n# Augmentation techniques\n# NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to add noise to embeddings\n# currently only supported on Llama and Mistral\nneftune_noise_alpha:\n\n# Optional[bool]. Whether to bettertransformers\nflash_optimum:\n\n# Note: Only one of the following attention patches can be used at a time.\n# For example, if you set `xformers_attention` to `true`, do not set `flash_attention` to `true`.\n\n# Optional[bool]. Whether to use xformers attention patch https://github.com/facebookresearch/xformers:\nxformers_attention:\n# Optional[bool]. Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:\nflash_attention:\nflash_attn_cross_entropy: # Optional[bool]. Whether to use flash-attention cross entropy implementation - advanced use only\nflash_attn_rms_norm: # Optional[bool]. Whether to use flash-attention rms norm implementation - advanced use only\nflash_attn_fuse_qkv: # Optional[bool]. Whether to fuse QKV into a single operation\nflash_attn_fuse_mlp: # Optional[bool]. Whether to fuse part of the MLP into a single operation\n# Optional[bool]. Whether to use scaled-dot-product attention\n# https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html\nsdp_attention:\n# Optional[bool]. Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf\ns2_attention:\n\n# Optional[bool]. Whether to use low_cpu_mem_usage\nlow_cpu_mem_usage:\n# Optional[str]. Resume from a specific checkpoint dir\nresume_from_checkpoint:\n# Optional[bool]. 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: false\n\n## Multimodal section\n# int | tuple[int, int] | None . Size to resize images to, width x height.\n# Will read from model/processor config if not set.\nimage_size:\n# str. Algorithm to use for image resizing. \"bilinear\", \"bicubic\", \"lanczos\". Default is \"bilinear\".\nimage_resize_algorithm: 'bilinear'\n## End of multimodal section\n\n# Don't mess with this, it's here for accelerate and torchrun\nlocal_rank:\n\n# Add or change special tokens.\n# If you add tokens here, you don't need to add them to the `tokens` list.\nspecial_tokens:\n # bos_token: \"<s>\"\n # eos_token: \"</s>\"\n # unk_token: \"<unk>\"\n # pad_token: \"[PAD]\"\n\n# Optional[list[str]]. Add extra tokens to the tokenizer.\ntokens:\n # - \"<|startoftext|>\"\n # - \"<|endoftext|>\"\n\n# Mapping token_id to new_token_string to override reserved added_tokens in the tokenizer.\n# Only works for tokens that are not part of the base vocab (aka are added_tokens).\n# Can be checked if they exist in tokenizer.json added_tokens.\nadded_tokens_overrides: # Dict[int, str]\n# 128041: \"<|im_start|>\"\n# 128042: \"<|im_end|>\"\n\n# FSDP\nfsdp:\nfsdp_config:\n\n# Deepspeed config path. e.g., deepspeed_configs/zero3.json\ndeepspeed:\n\n# Advanced DDP Arguments\nddp_timeout:\nddp_bucket_cap_mb:\nddp_broadcast_buffers:\n\n# Sequence parallelism\n# Set to a divisor of the number of GPUs available to split sequences into chunks of equal size.\n# Use in long context training to prevent OOM when sequences cannot fit into a single GPU's VRAM.\n# E.g., if 4 GPUs are available, set this value to 2 to split each sequence into two equal-sized\n# subsequences, or set to 4 to split into four equal-sized subsequences.\n# See https://docs.axolotl.ai/docs/sequence_parallelism.html for more details.\nsequence_parallel_degree:\n# Optional; strides across the key dimension. Larger values use more memory but should make training faster.\n# Must evenly divide the number of KV heads in your model.\nheads_k_stride: 1\n# One of \"varlen_llama3\", \"batch_ring\", \"batch_zigzag\", \"batch_stripe\". Defaults to \"varlen_llama3\"\n# in the sample packing case, and \"batch_ring\" in the non-sample packing case.\nring_attn_func:\n\n# Path to torch distx for optim 'adamw_anyprecision'\ntorchdistx_path:\n\n# Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize\npretraining_dataset:\n\n# Debug mode\ndebug:\n\n# Seed\nseed:\n\n# Allow overwrite yml config using from cli\nstrict:", + "text": "# This is the huggingface model that contains *.pt, *.safetensors, or *.bin files\n# This can also be a relative path to a model on disk\nbase_model: ./llama-7b-hf\n# You can specify an ignore pattern if the model repo contains more than 1 model type (*.pt, etc)\nbase_model_ignore_patterns:\n# If the base_model repo on hf hub doesn't include configuration .json files,\n# You can set that here, or leave this empty to default to base_model\nbase_model_config: ./llama-7b-hf\n# You can specify to choose a specific model revision from huggingface hub\nrevision_of_model:\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:\n# If you want to specify the type of model to load, AutoModelForCausalLM is a good choice too\nmodel_type: AutoModelForCausalLM\n# Corresponding tokenizer for the model AutoTokenizer is a good choice\ntokenizer_type: AutoTokenizer\n# Trust remote code for untrusted source\ntrust_remote_code:\n# use_fast option for tokenizer loading from_pretrained, default to True\ntokenizer_use_fast:\n# Whether to use the legacy tokenizer setting, defaults to True\ntokenizer_legacy:\n# Resize the model embeddings when new tokens are added to multiples of 32\n# This is reported to improve training speed on some models\nresize_token_embeddings_to_32x:\n# Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.\nshrink_embeddings:\n# Whether to load the model with randomly initialized weights. Useful for\n# pre-training a model from scratch or debugging purposes.\nrandom_init_weights:\n\n# (Internal use only)\n# Used to identify which the model is based on\nis_falcon_derived_model:\nis_llama_derived_model:\nis_qwen_derived_model:\n# Please note that if you set this to true, `padding_side` will be set to \"left\" by default\nis_mistral_derived_model:\n\n# optional overrides to the base model configuration\noverrides_of_model_config:\n # RoPE Scaling https://github.com/huggingface/transformers/pull/24653\n rope_scaling:\n type: # linear | dynamic\n factor: # float\n\n# optional overrides the base model loading from_pretrained\noverrides_of_model_kwargs:\n # use_cache: False\n\n# optional overrides to the bnb 4bit quantization configuration\n# https://huggingface.co/docs/transformers/main/main_classes/quantization#transformers.BitsAndBytesConfig\nbnb_config_kwargs:\n # These are default values\n llm_int8_has_fp16_weight: false\n bnb_4bit_quant_type: nf4\n bnb_4bit_use_double_quant: true\n\n\n# Whether you are training a 4-bit GPTQ quantized model\ngptq: true\n\n# This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer\nload_in_8bit: true\n# Use bitsandbytes 4 bit\nload_in_4bit:\n\n# Use CUDA bf16\nbf16: true # bool or 'full' for `bf16_full_eval`. require >=ampere\n# Use CUDA fp16\nfp16: true\n# Use CUDA tf32\ntf32: true # require >=ampere\n\n# No AMP (automatic mixed precision)\nbfloat16: true # require >=ampere\nfloat16: true\n\n# Limit the memory for all available GPUs to this amount (if an integer, expressed in gigabytes); default: unset\ngpu_memory_limit: 20GiB\n# 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\nlora_on_cpu: true\n\n# List[str]. Add plugins to extend the pipeline.\n# See `src/axolotl/integrations` for the available plugins or doc below for more details.\n# https://docs.axolotl.ai/docs/custom_integrations.html\nplugins:\n # - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin\n\n# A list of one or more datasets to finetune the model with\ndatasets:\n # HuggingFace dataset repo | s3://,gs:// path | \"json\" for local dataset, make sure to fill data_files\n - path: vicgalle/alpaca-gpt4\n # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]\n type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>\n ds_type: # Optional[str] (json|arrow|parquet|text|csv) defines the datatype when path is a file\n data_files: # Optional[str] path to source data files\n\n shards: # Optional[int] split dataset into N pieces (use with shards_idx)\n shards_idx: # Optional[int] = 0 the index of sharded dataset to use\n\n preprocess_shards: # Optional[int] process dataset in N sequential chunks for memory efficiency (exclusive with `shards`)\n\n name: # Optional[str] name of dataset configuration to load\n split: train # Optional[str] name of dataset split to load from\n revision: # Optional[str] 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.\n trust_remote_code: # Optional[bool] Trust remote code for untrusted source\n\n # Custom user instruction prompt\n - path: repo\n type:\n # The below are defaults. only set what's needed if you use a different column name.\n system_prompt: \"\"\n system_format: \"{system}\"\n field_system: system\n field_instruction: instruction\n field_input: input\n field_output: output\n\n # Customizable to be single line or multi-line\n # Use {instruction}/{input} as key to be replaced\n # 'format' can include {input}\n format: |-\n User: {instruction} {input}\n Assistant:\n # 'no_input_format' cannot include {input}\n no_input_format: \"{instruction} \"\n\n # For `completion` datsets only, uses the provided field instead of `text` column\n field:\n\n # Using chat template\n - path: ...\n # Set type to `chat_template` to use this strategy\n type: chat_template\n # Specify the name of the chat template to use\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 tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default.\n # - 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\n # - 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.\n # - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.\n chat_template: tokenizer_default\n\n # Custom jinja chat template. Used only if `chat_template: jinja` or empty.\n chat_template_jinja:\n\n # Key containing the messages (default: \"messages\")\n field_messages: messages\n\n # Key containing the system message (default: \"system\")\n # If the system message is not present in the dataset sample, it will be loaded from the field_system property.\n field_system: system\n\n # Mapping of properties from the input dataset to the chat template.\n # (default: message_property_mappings={'role':'role', 'content':'content'})\n # If a property exists in the template but not in this mapping, the system will attempt\n # to load it directly from the message using the property name as the key.\n # Example: In the mapping below, 'from' is loaded from input dataset and used as 'role',\n # while 'value' is loaded and used as 'content' in the chat template.\n message_property_mappings:\n role: from\n content: value\n # ...\n\n # Optional[Dict[str, List]]. Roles mapping in the messages.\n # The format is {target_role: [source_roles]}. All source roles will be mapped to the target role.\n # The default is:\n roles:\n user: [\"human\", \"user\"]\n assistant: [\"gpt\", \"assistant\"]\n system: [\"system\"]\n tool: [\"tool\"]\n\n # Optional[bool]. 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 you wish to,\n # we recommend using a custom jinja template with the default system message removed or\n # adding a system turn with empty content.\n drop_system_message:\n\n # Optional[bool]. Whether to split the assistant turn based on a reasoning trace inside delimited tags\n # defaults to False\n split_thinking:\n\n # IMPORTANT: The following fields determine which parts of the conversation to train on.\n # Priority order: message_field_training > message_field_training_detail > train_on_inputs or role in roles_to_train\n # See examples at `docs/dataset-formats/conversation.qmd`\n # Note: If the below 5 fields are empty, defaults to training only on the last message.\n\n # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.\n roles_to_train: [\"assistant\"] # default\n # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are:\n # - all: train on all EOS tokens\n # - turn (default): train on the EOS token at the end of each trainable turn\n # - last: train on the last EOS token in the conversation\n # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.\n train_on_eos: turn\n # Optional[str]. Which EOT (End-of-Turn) tokens to train on in the conversation. Possible values are:\n # - all: train on all EOT tokens\n # - turn: train on the EOT token at the end of each trainable turn\n # - last: train on the last EOT token in the conversation\n # If not specified, defaults to the value of train_on_eos for backward compatibility.\n train_on_eot:\n # 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`.\n message_field_training: training\n # The key in the message turn that contains the training details. Useful to selectively train on certain tokens in a turn.\n # 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).\n message_field_training_detail: train_detail\n\n\n# If false, the datasets will not be shuffled and will keep their original order in `datasets`.\n# The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.\nshuffle_merged_datasets: true\n\nDeduplicates datasets and test_datasets with identical entries.\ndataset_exact_deduplication: true\n\n# A list of one or more datasets to eval the model with.\n# You can use either test_datasets, or val_set_size, but not both.\ntest_datasets:\n - path: /workspace/data/eval.jsonl\n ds_type: json\n # You need to specify a split. For \"json\" datasets the default split is called \"train\".\n split: train\n type: completion\n data_files:\n - /workspace/data/eval.jsonl\n\n# use RL training: 'dpo', 'ipo', 'kto', 'simpo', 'orpo', 'grpo'\nrl:\nrl_beta: # Optional[float]. The beta parameter for the RL training.\n\n# dpo\ndpo_use_weighting: # Optional[bool]. Whether to perform weighting.\nrpo_alpha: # Optional[float]. Weighting of NLL term in loss from RPO paper.\n\n# orpo\norpo_alpha: 0.1 # Parameter controlling the relative ratio loss weight in the ORPO loss. Passed to `beta` in `ORPOConfig` due to trl mapping.\n\n# kto\nkto_desirable_weight: # Optional[float]. Factor for desirable loss term in KTO loss.\nkto_undesirable_weight: # Optional[float]. Factor for undesirable loss term in KTO loss.\n\n# simpo\ncpo_alpha: 1.0 # Weight of the BC regularizer\nsimpo_gamma: 0.5 # Target reward margin for the SimPO loss\n\n# grpo\ntrl:\n use_vllm: # Optional[bool]. Whether to use VLLM for RL training.\n vllm_server_host: # Optional[str]. Host of the vLLM server to connect to.\n vllm_server_port: # Optional[int]. Port of the vLLM server to connect to.\n vllm_server_timeout: # Optional[int]. Total timeout (in seconds) to wait for the vLLM server to respond.\n vllm_guided_decoding_regex: # Optional[str]. Regex for vLLM guided decoding.\n\n beta: # Optional[float]. Beta parameter for the RL training. Same as `rl_beta`. Use\n max_completion_length: # Optional[int]. Maximum length of the completion for RL training.\n\n reward_funcs: # Optional[list[str]]. List of reward functions to load. Paths must be importable from current dir.\n reward_weights: # Optional[list[float]]. List of reward weights for the reward functions.\n\n num_generations: # Optional[int]. Number of generations to sample.\n log_completions: # Optional[bool]. Whether to log completions.\n\n sync_ref_model: # Optional[bool]. Whether to sync the reference model.\n ref_model_mixup_alpha: # Optional[float]. Mixup alpha for the reference model.\n ref_model_sync_steps: # Optional[int]. Sync steps for the reference model.\n\n\n# reward modelling: `True` or `False`\nreward_model:\n\n# process reward modelling: `True` or `False`\nprocess_reward_model:\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 tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default value.\n# - 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\n# - 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.\n# - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.\n# The selected chat template will be saved to the tokenizer_config.json for easier inferencing\n# Note: It is recommended to set train_on_inputs to true when using a chat template that is different from the model's default chat template.\nchat_template: tokenizer_default\n# custom jinja template 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.\nchat_template_jinja: null\n# Optional[List[str]]. Custom EOT (End-of-Turn) tokens to mask/unmask during training.\n# These tokens mark the boundaries between conversation turns.\n# For example: [\"/INST\", \"</s>\", \"[/SYSTEM_PROMPT]\"]\n# If not specified, defaults to just the model's eos_token.\n# This is useful for templates that use multiple delimiter tokens.\neot_tokens:\n # - \"</s>\"\n # - \"[/INST]\"\n # - \"[/SYSTEM_PROMPT]\"\n# Changes the default system message\ndefault_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml.\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: data/last_run_prepared\n# Push prepared dataset to hub\npush_dataset_to_hub: # Optional[str] repo_org/repo_name\n# The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`\n# if not set.\ndataset_processes: # defaults to os.cpu_count() if not set\n# Keep dataset in memory while preprocessing\n# Only needed if cached dataset is taking too much storage\ndataset_keep_in_memory:\n# push checkpoints to hub\nhub_model_id: # private repo path to push finetuned model\n# how to push checkpoints to hub\n# https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy\nhub_strategy:\n# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets\n# Required to be true when used in combination with `push_dataset_to_hub`\nhf_use_auth_token: # boolean\n# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval.\nval_set_size: 0.04\n# Num shards for whole dataset\ndataset_shard_num:\n# Index of shard to use for whole dataset\ndataset_shard_idx:\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: 2048\n# Pad inputs so each step uses constant sized buffers\n# This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently\npad_to_sequence_len:\n# Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'\nsample_packing:\n# Set to 'false' if getting errors during eval with sample_packing on.\neval_sample_packing:\n# You can set these packing optimizations AFTER starting a training at least once.\n# The trainer will provide recommended values for these values.\nsample_packing_eff_est:\ntotal_num_tokens:\n# Increasing the following values helps with packing, but usually only slightly (<%1.)\n# The number of samples packed at a time.\nsample_packing_group_size: 100000\n# The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.\nsample_packing_bin_size: 200\nsample_pack_sequentially: # Optional[bool]. Whether to pack samples sequentially.\n\n# whether to concatenate samples during pretraining\npretraining_sample_concatenation:\n\ncurriculum_sampling: # Optional[bool]. Whether to use sequential sampling for curriculum learning\n\n# Use batch flattening for speedups when not using sample_packing\nbatch_flattening:\n\n# Passed through to transformers when loading the model when launched without accelerate\n# Use `sequential` when training w/ model parallelism to limit memory\ndevice_map:\n# Defines the max memory usage per gpu on the system. Passed through to transformers when loading the model.\nmax_memory:\n\n# If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model\nadapter: lora\n# If you already have a lora model trained that you want to load, put that here.\n# This means after training, if you want to test the model, you should set this to the value of `output_dir`.\n# Note that if you merge an adapter to the base model, a new subdirectory `merged` will be created under the `output_dir`.\nlora_model_dir:\n\n# LoRA hyperparameters\n# For more details about the following options, see:\n# https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2\nlora_r: 8\nlora_alpha: 16\nlora_dropout: 0.05\nlora_target_modules:\n - q_proj\n - v_proj\n# - k_proj\n# - o_proj\n# - gate_proj\n# - down_proj\n# - up_proj\nlora_target_linear: # If true, will target all linear modules\n\n# List[int] | int. # The layer indices to transform, otherwise, apply to all layers\n# https://huggingface.co/docs/peft/v0.15.0/en/package_reference/lora#peft.LoraConfig.layers_to_transform\npeft_layers_to_transform:\n\n# Optional[bool]. Whether to use DoRA.\n# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#weight-decomposed-low-rank-adaptation-dora\npeft_use_dora:\n\n# Optional[bool]. Whether to use RSLoRA.\n# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#rank-stabilized-lora\npeft_use_rslora:\n\n# Optional[list[tuple[int, int]]]. List of layer indices to replicate.\n# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#memory-efficient-layer-replication-with-lora\npeft_layer_replication:\n\n# bool | Literal[\"gaussian\", \"eva\", \"olora\", \"pissa\", \"pissa_niter_[number of iters]\", \"corda\", \"loftq\"]\n# How to initialize LoRA weights. Default to True which is MS original implementation.\n# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#initialization\npeft_init_lora_weights:\n\n# If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.\n# For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.\n# `embed_tokens` converts tokens to embeddings, and `lm_head` converts embeddings to token probabilities.\n# https://github.com/huggingface/peft/issues/334#issuecomment-1561727994\nlora_modules_to_save:\n# - embed_tokens\n# - lm_head\n\nlora_fan_in_fan_out: false\n\n# Apply custom LoRA autograd functions and activation function Triton kernels for\n# speed and memory savings\n# See: https://docs.axolotl.ai/docs/lora_optims.html\nlora_mlp_kernel: true\nlora_qkv_kernel: true\nlora_o_kernel: true\n\n# LoRA+ hyperparameters\n# For more details about the following options, see:\n# https://arxiv.org/abs/2402.12354 and `src/axolotl/core/train_builder.py`\nloraplus_lr_ratio: # loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4.\nloraplus_lr_embedding: # loraplus learning rate for lora embedding layers. Default value is 1e-6.\n\npeft:\n # Configuration options for loftq initialization for LoRA\n # https://huggingface.co/docs/peft/developer_guides/quantization#loftq-initialization\n loftq_config:\n loftq_bits: # typically 4 bits\n\n# ReLoRA configuration\n# Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed\nrelora_steps: # Number of steps per ReLoRA restart\nrelora_warmup_steps: # Number of per-restart warmup steps\nrelora_anneal_steps: # Number of anneal steps for each relora cycle\nrelora_prune_ratio: # threshold for optimizer magnitude when pruning\nrelora_cpu_offload: # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings\n\n# wandb configuration if you're using it\n# Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.\nwandb_mode: # \"offline\" to save run metadata locally and not sync to the server, \"disabled\" to turn off wandb\nwandb_project: # Your wandb project name\nwandb_entity: # A wandb Team name if using a Team\nwandb_watch:\nwandb_name: # Set the name of your wandb run\nwandb_run_id: # Set the ID of your wandb run\nwandb_log_model: # \"checkpoint\" to log model to wandb Artifacts every `save_steps` or \"end\" to log only at the end of training\n\n# mlflow configuration if you're using it\nmlflow_tracking_uri: # URI to mlflow\nmlflow_experiment_name: # Your experiment name\nmlflow_run_name: # Your run name\nhf_mlflow_log_artifacts: # set to true to copy each saved checkpoint on each save to mlflow artifact registry\n\n# Comet configuration if you're using it\n# Make sure your `COMET_API_KEY` environment variable is set (recommended) or you login to Comet with `comet login`.\n# Check out our documentation for more details https://www.comet.com/docs/v2/api-and-sdk/python-sdk/reference/Experiment-Creation/#comet_ml.start\nuse_comet: # Enable or disable Comet integration.\ncomet_api_key: # API key for Comet. Recommended to set via `comet login`.\ncomet_workspace: # Workspace name in Comet. Defaults to the user's default workspace.\ncomet_project_name: # Project name in Comet. Defaults to Uncategorized.\ncomet_experiment_key: # Identifier for the experiment. Used to append data to an existing experiment or control the key of new experiments. Default to a random key.\ncomet_mode: # Create a new experiment (\"create\") or log to an existing one (\"get\"). Default (\"get_or_create\") auto-selects based on configuration.\ncomet_online: # Set to True to log data to Comet server, or False for offline storage. Default is True.\ncomet_experiment_config: # Dictionary for additional configuration settings, see the doc for more details.\n\n# Tensorboard\nuse_tensorboard: # Optional[bool]\n\n# Where to save the full-finetuned model to\noutput_dir: ./completed-model\n\n# Whether to use torch.compile and which backend to use\n# setting to `auto` will enable torch compile when torch>=2.5.1\ntorch_compile: # Optional[Union[Literal[\"auto\"], bool]]\ntorch_compile_backend: # Optional[str]\n\n# Training hyperparameters\n\n# If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps.\ngradient_accumulation_steps: 1\n# The number of samples to include in each batch. This is the number of samples sent to each GPU.\n# Batch size per gpu = micro_batch_size * gradient_accumulation_steps\nmicro_batch_size: 2\neval_batch_size:\nnum_epochs: 4\nwarmup_steps: 100 # cannot use with warmup_ratio\nwarmup_ratio: 0.05 # cannot use with warmup_steps\nlearning_rate: 0.00003\nlr_quadratic_warmup:\nlogging_steps:\neval_steps: # Leave empty to eval at each epoch, integer for every N steps. float for fraction of total steps\nevals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps\neval_strategy: # Set to `\"no\"` to skip evaluation, `\"epoch\"` at end of each epoch, leave empty to infer from `eval_steps`.\nsave_strategy: # 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`.\nsave_steps: # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps\nsaves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps\nsave_total_limit: # Checkpoints saved at a time\n# Maximum number of iterations to train for. It precedes num_epochs which means that\n# if both are set, num_epochs will not be guaranteed.\n# e.g., when 1 epoch is 1000 steps => `num_epochs: 2` and `max_steps: 100` will train for 100 steps\nmax_steps:\n\n# 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.\ninclude_tokens_per_second: # Optional[bool]\n\n# whether to find batch size that fits in memory. Passed to underlying transformers Trainer\nauto_find_batch_size: # Optional[bool]\n\neval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0\neval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128\ndo_causal_lm_eval: # Whether to run causal language model evaluation for metrics in `eval_causal_lm_metrics`.\neval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is [\"sacrebleu\", \"comet\", \"ter\", \"chrf\", \"perplexity\"]\n\nprofiler_steps: # enable the pytorch profiler to capture the first N steps of training to the output_dir.\n # see https://pytorch.org/blog/understanding-gpu-memory-1/ for more information\n # snapshots can be visualized @ https://pytorch.org/memory_viz\n\nloss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)\nloss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3)\n\n# Save model as safetensors (require safetensors package)\nsave_safetensors:\n\n# Whether to mask out or include the human's prompt from the training labels\ntrain_on_inputs: false\n# Group similarly sized data to minimize padding.\n# May be slower to start, as it must download and sort the entire dataset.\n# Note that training loss may have an oscillating pattern with this enabled.\ngroup_by_length: false\n\n# Whether to use gradient checkpointing. Available options are: true, false, \"offload\".\n# https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing\ngradient_checkpointing: false\n# additional kwargs to pass to the trainer for gradient checkpointing\n# gradient_checkpointing_kwargs:\n# use_reentrant: true\n\n# Stop training after this many evaluation losses have increased in a row\n# https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback\nearly_stopping_patience: 3\n\n# Specify a scheduler and kwargs to use with the optimizer\nlr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | empty for cosine\nlr_scheduler_kwargs:\ncosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr\ncosine_constant_lr_ratio: # 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 (https://arxiv.org/pdf/2308.04014.pdf)\n\n# For one_cycle optim\nlr_div_factor: # Learning rate div factor\n\n# Specify optimizer\n# Valid values are driven by the Transformers OptimizerNames class, see:\n# https://github.com/huggingface/transformers/blob/cbf924b76c03828101a34069a96d209314114fd5/src/transformers/training_args.py#L144-L189\n#\n# Note that not all optimizers may be available in your environment, ex: 'adamw_anyprecision' is part of\n# torchdistx, 'adamw_bnb_8bit' is part of bnb.optim.Adam8bit, etc. When in doubt, it is recommended to start with the optimizer used\n# in the examples/ for your model and fine-tuning use case.\n#\n# Valid values for 'optimizer' include:\n# - adamw_torch\n# - adamw_torch_fused\n# - adamw_torch_xla\n# - adamw_torch_npu_fused\n# - adamw_apex_fused\n# - adopt_adamw (an EXPERIMENTAL optimizer, only for torch version >= 2.5.1)\n# - adafactor\n# - adamw_anyprecision\n# - adamw_torch_4bit\n# - ademamix\n# - sgd\n# - adagrad\n# - adamw_bnb_8bit\n# - adamw_8bit # alias for adamw_bnb_8bit\n# - ademamix_8bit\n# - lion_8bit\n# - lion_32bit\n# - paged_adamw_32bit\n# - paged_adamw_8bit\n# - paged_ademamix_32bit\n# - paged_ademamix_8bit\n# - paged_lion_32bit\n# - paged_lion_8bit\n# - rmsprop\n# - rmsprop_bnb\n# - rmsprop_bnb_8bit\n# - rmsprop_bnb_32bit\n# - galore_adamw\n# - galore_adamw_8bit\n# - galore_adafactor\n# - galore_adamw_layerwise\n# - galore_adamw_8bit_layerwise\n# - galore_adafactor_layerwise\n# - lomo\n# - adalomo\n# - grokadamw\n# - schedule_free_adamw\n# - schedule_free_sgd\n# - apollo_adamw\n# - apollo_adamw_layerwise\n#\n# Additional custom optimizers include:\n# - optimi_adamw\n# - ao_adamw_8bit\n# - ao_adamw_fp8\noptimizer:\n# Dictionary of arguments to pass to the optimizer\noptim_args:\n# For Galore Optimizers the following optim_args are available\n# rank: # type: int\n# update_proj_gap # type: int\n# scale # type: float\n# proj_type: # type: str, default = std\n\n# 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\noptim_target_modules:\n# - self_attn # for llama\n# - mlp\n\n# Specify weight decay\nweight_decay:\n# adamw hyperparams\nadam_beta1:\nadam_beta2:\nadam_epsilon:\n# Gradient clipping max norm\nmax_grad_norm:\n\n# Augmentation techniques\n# NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to add noise to embeddings\n# currently only supported on Llama and Mistral\nneftune_noise_alpha:\n\n# Optional[bool]. Whether to bettertransformers\nflash_optimum:\n\n# Note: Only one of the following attention patches can be used at a time.\n# For example, if you set `xformers_attention` to `true`, do not set `flash_attention` to `true`.\n\n# Optional[bool]. Whether to use xformers attention patch https://github.com/facebookresearch/xformers:\nxformers_attention:\n# Optional[bool]. Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:\nflash_attention:\nflash_attn_cross_entropy: # Optional[bool]. Whether to use flash-attention cross entropy implementation - advanced use only\nflash_attn_rms_norm: # Optional[bool]. Whether to use flash-attention rms norm implementation - advanced use only\nflash_attn_fuse_qkv: # Optional[bool]. Whether to fuse QKV into a single operation\nflash_attn_fuse_mlp: # Optional[bool]. Whether to fuse part of the MLP into a single operation\n# Optional[bool]. Whether to use scaled-dot-product attention\n# https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html\nsdp_attention:\n# Optional[bool]. Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf\ns2_attention:\n\n# Optional[bool]. Whether to use low_cpu_mem_usage\nlow_cpu_mem_usage:\n# Optional[str]. Resume from a specific checkpoint dir\nresume_from_checkpoint:\n# Optional[bool]. 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: false\n\n## Multimodal section\n# int | tuple[int, int] | None . Size to resize images to, width x height.\n# Will read from model/processor config if not set.\nimage_size:\n# str. Algorithm to use for image resizing. \"bilinear\", \"bicubic\", \"lanczos\". Default is \"bilinear\".\nimage_resize_algorithm: 'bilinear'\n## End of multimodal section\n\n# Don't mess with this, it's here for accelerate and torchrun\nlocal_rank:\n\n# Add or change special tokens.\n# If you add tokens here, you don't need to add them to the `tokens` list.\nspecial_tokens:\n # bos_token: \"<s>\"\n # eos_token: \"</s>\"\n # unk_token: \"<unk>\"\n # pad_token: \"[PAD]\"\n\n# Optional[list[str]]. Add extra tokens to the tokenizer.\ntokens:\n # - \"<|startoftext|>\"\n # - \"<|endoftext|>\"\n\n# Mapping token_id to new_token_string to override reserved added_tokens in the tokenizer.\n# Only works for tokens that are not part of the base vocab (aka are added_tokens).\n# Can be checked if they exist in tokenizer.json added_tokens.\nadded_tokens_overrides: # Dict[int, str]\n# 128041: \"<|im_start|>\"\n# 128042: \"<|im_end|>\"\n\n# FSDP\nfsdp:\nfsdp_config:\n\n# Deepspeed config path. e.g., deepspeed_configs/zero3.json\ndeepspeed:\n\n# Advanced DDP Arguments\nddp_timeout:\nddp_bucket_cap_mb:\nddp_broadcast_buffers:\n\n# Sequence parallelism\n# Set to a divisor of the number of GPUs available to split sequences into chunks of equal size.\n# Use in long context training to prevent OOM when sequences cannot fit into a single GPU's VRAM.\n# E.g., if 4 GPUs are available, set this value to 2 to split each sequence into two equal-sized\n# subsequences, or set to 4 to split into four equal-sized subsequences.\n# See https://docs.axolotl.ai/docs/sequence_parallelism.html for more details.\nsequence_parallel_degree:\n# Optional; strides across the key dimension. Larger values use more memory but should make training faster.\n# Must evenly divide the number of KV heads in your model.\nheads_k_stride: 1\n# One of \"varlen_llama3\", \"batch_ring\", \"batch_zigzag\", \"batch_stripe\". Defaults to \"varlen_llama3\"\n# in the sample packing case, and \"batch_ring\" in the non-sample packing case.\nring_attn_func:\n\n# Path to torch distx for optim 'adamw_anyprecision'\ntorchdistx_path:\n\n# Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize\npretraining_dataset:\n\n# Debug mode\ndebug:\n\n# Seed\nseed:\n\n# Allow overwrite yml config using from cli\nstrict:", "crumbs": [ "Getting Started", "Config Reference" diff --git a/sitemap.xml b/sitemap.xml index 7d8104321..45057e0e5 100644 --- a/sitemap.xml +++ b/sitemap.xml @@ -2,682 +2,682 @@ https://docs.axolotl.ai/examples/colab-notebooks/colab-axolotl-example.html - 2025-04-29T12:28:44.174Z + 2025-04-29T14:01:58.464Z https://docs.axolotl.ai/index.html - 2025-04-29T12:28:44.186Z + 2025-04-29T14:01:58.476Z https://docs.axolotl.ai/docs/rlhf.html - 2025-04-29T12:28:44.174Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/unsloth.html - 2025-04-29T12:28:44.174Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/dataset_preprocessing.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/input_output.html - 2025-04-29T12:28:44.173Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/dataset_loading.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/api/utils.collators.mamba.html - 2025-04-29T12:29:19.344Z + 2025-04-29T14:02:29.185Z https://docs.axolotl.ai/docs/api/utils.optimizers.adopt.html - 2025-04-29T12:29:19.122Z + 2025-04-29T14:02:28.961Z https://docs.axolotl.ai/docs/api/prompt_strategies.user_defined.html - 2025-04-29T12:29:18.637Z + 2025-04-29T14:02:28.472Z https://docs.axolotl.ai/docs/api/utils.chat_templates.html - 2025-04-29T12:29:19.024Z + 2025-04-29T14:02:28.864Z https://docs.axolotl.ai/docs/api/cli.merge_lora.html - 2025-04-29T12:29:18.467Z + 2025-04-29T14:02:28.301Z https://docs.axolotl.ai/docs/api/monkeypatch.multipack.html - 2025-04-29T12:29:18.909Z + 2025-04-29T14:02:28.745Z https://docs.axolotl.ai/docs/api/core.chat.format.shared.html - 2025-04-29T12:29:18.343Z + 2025-04-29T14:02:28.177Z https://docs.axolotl.ai/docs/api/utils.schemas.integrations.html - 2025-04-29T12:29:19.194Z + 2025-04-29T14:02:29.034Z https://docs.axolotl.ai/docs/api/utils.freeze.html - 2025-04-29T12:29:19.049Z + 2025-04-29T14:02:28.889Z https://docs.axolotl.ai/docs/api/prompt_strategies.alpaca_w_system.html - 2025-04-29T12:29:18.629Z + 2025-04-29T14:02:28.464Z https://docs.axolotl.ai/docs/api/monkeypatch.attention.mllama.html - 2025-04-29T12:29:18.976Z + 2025-04-29T14:02:28.815Z https://docs.axolotl.ai/docs/api/utils.schemas.model.html - 2025-04-29T12:29:19.143Z + 2025-04-29T14:02:28.982Z https://docs.axolotl.ai/docs/api/core.datasets.transforms.chat_builder.html - 2025-04-29T12:29:18.356Z + 2025-04-29T14:02:28.190Z https://docs.axolotl.ai/docs/api/monkeypatch.btlm_attn_hijack_flash.html - 2025-04-29T12:29:18.951Z + 2025-04-29T14:02:28.788Z https://docs.axolotl.ai/docs/api/models.mamba.modeling_mamba.html - 2025-04-29T12:29:19.320Z + 2025-04-29T14:02:29.161Z https://docs.axolotl.ai/docs/api/core.datasets.chat.html - 2025-04-29T12:29:18.348Z + 2025-04-29T14:02:28.182Z https://docs.axolotl.ai/docs/api/utils.model_shard_quant.html - 2025-04-29T12:29:19.038Z + 2025-04-29T14:02:28.878Z https://docs.axolotl.ai/docs/api/monkeypatch.llama_patch_multipack.html - 2025-04-29T12:29:18.953Z + 2025-04-29T14:02:28.789Z https://docs.axolotl.ai/docs/api/utils.lora_embeddings.html - 2025-04-29T12:29:19.033Z + 2025-04-29T14:02:28.872Z https://docs.axolotl.ai/docs/api/utils.schemas.peft.html - 2025-04-29T12:29:19.174Z + 2025-04-29T14:02:29.013Z https://docs.axolotl.ai/docs/api/core.chat.format.llama3x.html - 2025-04-29T12:29:18.342Z + 2025-04-29T14:02:28.175Z https://docs.axolotl.ai/docs/api/monkeypatch.mistral_attn_hijack_flash.html - 2025-04-29T12:29:18.907Z + 2025-04-29T14:02:28.743Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.chat_template.html - 2025-04-29T12:29:18.689Z + 2025-04-29T14:02:28.524Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.user_defined.html - 2025-04-29T12:29:18.712Z + 2025-04-29T14:02:28.548Z https://docs.axolotl.ai/docs/api/utils.distributed.html - 2025-04-29T12:29:19.111Z + 2025-04-29T14:02:28.950Z https://docs.axolotl.ai/docs/api/prompt_strategies.pygmalion.html - 2025-04-29T12:29:18.683Z + 2025-04-29T14:02:28.519Z https://docs.axolotl.ai/docs/api/utils.dict.html - 2025-04-29T12:29:19.114Z + 2025-04-29T14:02:28.954Z https://docs.axolotl.ai/docs/api/monkeypatch.llama_attn_hijack_flash.html - 2025-04-29T12:29:18.892Z + 2025-04-29T14:02:28.728Z https://docs.axolotl.ai/docs/api/prompt_strategies.base.html - 2025-04-29T12:29:18.583Z + 2025-04-29T14:02:28.418Z https://docs.axolotl.ai/docs/api/monkeypatch.data.batch_dataset_fetcher.html - 2025-04-29T12:29:18.978Z + 2025-04-29T14:02:28.817Z https://docs.axolotl.ai/docs/api/kernels.swiglu.html - 2025-04-29T12:29:18.857Z + 2025-04-29T14:02:28.693Z https://docs.axolotl.ai/docs/api/core.chat.messages.html - 2025-04-29T12:29:18.339Z + 2025-04-29T14:02:28.172Z https://docs.axolotl.ai/docs/api/index.html - 2025-04-29T12:29:18.076Z + 2025-04-29T14:02:27.907Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.zephyr.html - 2025-04-29T12:29:18.711Z + 2025-04-29T14:02:28.546Z https://docs.axolotl.ai/docs/api/convert.html - 2025-04-29T12:29:18.168Z + 2025-04-29T14:02:28.000Z https://docs.axolotl.ai/docs/api/utils.schemas.multimodal.html - 2025-04-29T12:29:19.182Z + 2025-04-29T14:02:29.022Z https://docs.axolotl.ai/docs/api/core.trainers.base.html - 2025-04-29T12:29:18.552Z + 2025-04-29T14:02:28.387Z https://docs.axolotl.ai/docs/api/evaluate.html - 2025-04-29T12:29:18.147Z + 2025-04-29T14:02:27.979Z https://docs.axolotl.ai/docs/api/monkeypatch.llama_attn_hijack_xformers.html - 2025-04-29T12:29:18.893Z + 2025-04-29T14:02:28.729Z https://docs.axolotl.ai/docs/api/kernels.quantize.html - 2025-04-29T12:29:18.864Z + 2025-04-29T14:02:28.701Z https://docs.axolotl.ai/docs/api/utils.callbacks.mlflow_.html - 2025-04-29T12:29:19.372Z + 2025-04-29T14:02:29.213Z https://docs.axolotl.ai/docs/api/utils.callbacks.profiler.html - 2025-04-29T12:29:19.367Z + 2025-04-29T14:02:29.208Z https://docs.axolotl.ai/docs/api/core.trainers.dpo.trainer.html - 2025-04-29T12:29:18.579Z + 2025-04-29T14:02:28.413Z https://docs.axolotl.ai/docs/api/cli.vllm_serve.html - 2025-04-29T12:29:18.528Z + 2025-04-29T14:02:28.363Z https://docs.axolotl.ai/docs/api/train.html - 2025-04-29T12:29:18.136Z + 2025-04-29T14:02:27.969Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.chatml.html - 2025-04-29T12:29:18.709Z + 2025-04-29T14:02:28.545Z https://docs.axolotl.ai/docs/api/utils.schemas.trl.html - 2025-04-29T12:29:19.177Z + 2025-04-29T14:02:29.016Z https://docs.axolotl.ai/docs/api/kernels.geglu.html - 2025-04-29T12:29:18.847Z + 2025-04-29T14:02:28.683Z https://docs.axolotl.ai/docs/api/utils.bench.html - 2025-04-29T12:29:19.041Z + 2025-04-29T14:02:28.881Z https://docs.axolotl.ai/docs/api/monkeypatch.transformers_fa_utils.html - 2025-04-29T12:29:18.968Z + 2025-04-29T14:02:28.807Z https://docs.axolotl.ai/docs/api/integrations.liger.args.html - 2025-04-29T12:29:19.293Z + 2025-04-29T14:02:29.134Z https://docs.axolotl.ai/docs/api/core.trainer_builder.html - 2025-04-29T12:29:18.229Z + 2025-04-29T14:02:28.062Z https://docs.axolotl.ai/docs/api/utils.schemas.utils.html - 2025-04-29T12:29:19.206Z + 2025-04-29T14:02:29.046Z https://docs.axolotl.ai/docs/api/kernels.lora.html - 2025-04-29T12:29:18.837Z + 2025-04-29T14:02:28.673Z https://docs.axolotl.ai/docs/api/prompt_strategies.bradley_terry.llama3.html - 2025-04-29T12:29:18.755Z + 2025-04-29T14:02:28.591Z https://docs.axolotl.ai/docs/api/core.trainers.grpo.trainer.html - 2025-04-29T12:29:18.582Z + 2025-04-29T14:02:28.417Z https://docs.axolotl.ai/docs/api/prompt_strategies.messages.chat.html - 2025-04-29T12:29:18.688Z + 2025-04-29T14:02:28.523Z https://docs.axolotl.ai/docs/api/utils.tokenization.html - 2025-04-29T12:29:19.015Z + 2025-04-29T14:02:28.854Z https://docs.axolotl.ai/docs/api/prompt_strategies.chat_template.html - 2025-04-29T12:29:18.601Z + 2025-04-29T14:02:28.436Z https://docs.axolotl.ai/docs/api/prompt_strategies.stepwise_supervised.html - 2025-04-29T12:29:18.666Z + 2025-04-29T14:02:28.501Z https://docs.axolotl.ai/docs/api/utils.samplers.multipack.html - 2025-04-29T12:29:19.357Z + 2025-04-29T14:02:29.198Z https://docs.axolotl.ai/docs/api/cli.args.html - 2025-04-29T12:29:18.420Z + 2025-04-29T14:02:28.254Z https://docs.axolotl.ai/docs/api/utils.callbacks.perplexity.html - 2025-04-29T12:29:19.363Z + 2025-04-29T14:02:29.204Z https://docs.axolotl.ai/docs/api/utils.gradient_checkpointing.unsloth.html - 2025-04-29T12:29:19.128Z + 2025-04-29T14:02:28.967Z https://docs.axolotl.ai/docs/mac.html - 2025-04-29T12:28:44.173Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/config.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.459Z https://docs.axolotl.ai/docs/multimodal.html - 2025-04-29T12:28:44.173Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/lr_groups.html - 2025-04-29T12:28:44.173Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/dataset-formats/index.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/dataset-formats/stepwise_supervised.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/dataset-formats/pretraining.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/multi-node.html - 2025-04-29T12:28:44.173Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/sequence_parallelism.html - 2025-04-29T12:28:44.174Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/batch_vs_grad.html - 2025-04-29T12:28:44.169Z + 2025-04-29T14:01:58.459Z https://docs.axolotl.ai/docs/amd_hpc.html - 2025-04-29T12:28:44.169Z + 2025-04-29T14:01:58.459Z https://docs.axolotl.ai/docs/faq.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/custom_integrations.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.459Z https://docs.axolotl.ai/src/axolotl/integrations/LICENSE.html - 2025-04-29T12:28:44.190Z + 2025-04-29T14:01:58.480Z https://docs.axolotl.ai/TODO.html - 2025-04-29T12:28:44.168Z + 2025-04-29T14:01:58.458Z https://docs.axolotl.ai/src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.html - 2025-04-29T12:28:44.190Z + 2025-04-29T14:01:58.480Z https://docs.axolotl.ai/docs/getting-started.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/multipack.html - 2025-04-29T12:28:44.173Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/multi-gpu.html - 2025-04-29T12:28:44.173Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/installation.html - 2025-04-29T12:28:44.173Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/cli.html - 2025-04-29T12:28:44.169Z + 2025-04-29T14:01:58.459Z https://docs.axolotl.ai/docs/dataset-formats/inst_tune.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/dataset-formats/tokenized.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/dataset-formats/conversation.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/dataset-formats/template_free.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/reward_modelling.html - 2025-04-29T12:28:44.174Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/lora_optims.html - 2025-04-29T12:28:44.173Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/nccl.html - 2025-04-29T12:28:44.173Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/api/logging_config.html - 2025-04-29T12:29:18.214Z + 2025-04-29T14:02:28.047Z https://docs.axolotl.ai/docs/api/utils.trainer.html - 2025-04-29T12:29:19.066Z + 2025-04-29T14:02:28.906Z https://docs.axolotl.ai/docs/api/monkeypatch.unsloth_.html - 2025-04-29T12:29:18.970Z + 2025-04-29T14:02:28.808Z https://docs.axolotl.ai/docs/api/cli.evaluate.html - 2025-04-29T12:29:18.404Z + 2025-04-29T14:02:28.237Z https://docs.axolotl.ai/docs/api/kernels.utils.html - 2025-04-29T12:29:18.866Z + 2025-04-29T14:02:28.702Z https://docs.axolotl.ai/docs/api/datasets.html - 2025-04-29T12:29:18.154Z + 2025-04-29T14:02:27.987Z https://docs.axolotl.ai/docs/api/utils.models.html - 2025-04-29T12:29:19.008Z + 2025-04-29T14:02:28.847Z https://docs.axolotl.ai/docs/api/prompt_strategies.kto.llama3.html - 2025-04-29T12:29:18.722Z + 2025-04-29T14:02:28.557Z https://docs.axolotl.ai/docs/api/cli.preprocess.html - 2025-04-29T12:29:18.487Z + 2025-04-29T14:02:28.321Z https://docs.axolotl.ai/docs/api/cli.merge_sharded_fsdp_weights.html - 2025-04-29T12:29:18.479Z + 2025-04-29T14:02:28.313Z https://docs.axolotl.ai/docs/api/integrations.base.html - 2025-04-29T12:29:19.278Z + 2025-04-29T14:02:29.118Z https://docs.axolotl.ai/docs/api/prompt_strategies.orpo.chat_template.html - 2025-04-29T12:29:18.751Z + 2025-04-29T14:02:28.587Z https://docs.axolotl.ai/docs/api/utils.schemas.enums.html - 2025-04-29T12:29:19.201Z + 2025-04-29T14:02:29.041Z https://docs.axolotl.ai/docs/api/utils.callbacks.comet_.html - 2025-04-29T12:29:19.376Z + 2025-04-29T14:02:29.216Z https://docs.axolotl.ai/docs/api/prompt_strategies.input_output.html - 2025-04-29T12:29:18.661Z + 2025-04-29T14:02:28.497Z https://docs.axolotl.ai/docs/api/utils.schedulers.html - 2025-04-29T12:29:19.090Z + 2025-04-29T14:02:28.930Z https://docs.axolotl.ai/docs/api/cli.cloud.base.html - 2025-04-29T12:29:18.532Z + 2025-04-29T14:02:28.366Z https://docs.axolotl.ai/docs/api/cli.utils.html - 2025-04-29T12:29:18.524Z + 2025-04-29T14:02:28.358Z https://docs.axolotl.ai/docs/api/monkeypatch.lora_kernels.html - 2025-04-29T12:29:18.942Z + 2025-04-29T14:02:28.778Z https://docs.axolotl.ai/docs/api/prompt_strategies.completion.html - 2025-04-29T12:29:18.655Z + 2025-04-29T14:02:28.491Z https://docs.axolotl.ai/docs/api/utils.schemas.config.html - 2025-04-29T12:29:19.136Z + 2025-04-29T14:02:28.975Z https://docs.axolotl.ai/docs/api/monkeypatch.llama_expand_mask.html - 2025-04-29T12:29:18.917Z + 2025-04-29T14:02:28.753Z https://docs.axolotl.ai/docs/api/cli.config.html - 2025-04-29T12:29:18.444Z + 2025-04-29T14:02:28.279Z https://docs.axolotl.ai/docs/api/utils.collators.core.html - 2025-04-29T12:29:19.322Z + 2025-04-29T14:02:29.162Z https://docs.axolotl.ai/docs/api/utils.data.sft.html - 2025-04-29T12:29:19.125Z + 2025-04-29T14:02:28.964Z https://docs.axolotl.ai/docs/api/integrations.spectrum.args.html - 2025-04-29T12:29:19.300Z + 2025-04-29T14:02:29.140Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.llama3.html - 2025-04-29T12:29:18.699Z + 2025-04-29T14:02:28.535Z https://docs.axolotl.ai/docs/api/cli.inference.html - 2025-04-29T12:29:18.458Z + 2025-04-29T14:02:28.293Z https://docs.axolotl.ai/docs/api/prompt_strategies.alpaca_instruct.html - 2025-04-29T12:29:18.617Z + 2025-04-29T14:02:28.451Z https://docs.axolotl.ai/docs/api/utils.collators.batching.html - 2025-04-29T12:29:19.340Z + 2025-04-29T14:02:29.181Z https://docs.axolotl.ai/docs/api/utils.schemas.datasets.html - 2025-04-29T12:29:19.165Z + 2025-04-29T14:02:29.005Z https://docs.axolotl.ai/docs/api/utils.data.pretraining.html - 2025-04-29T12:29:19.124Z + 2025-04-29T14:02:28.963Z https://docs.axolotl.ai/docs/api/prompt_strategies.dpo.passthrough.html - 2025-04-29T12:29:18.714Z + 2025-04-29T14:02:28.549Z https://docs.axolotl.ai/docs/api/monkeypatch.utils.html - 2025-04-29T12:29:18.950Z + 2025-04-29T14:02:28.786Z https://docs.axolotl.ai/docs/api/utils.callbacks.lisa.html - 2025-04-29T12:29:19.368Z + 2025-04-29T14:02:29.209Z https://docs.axolotl.ai/docs/api/prompt_strategies.alpaca_chat.html - 2025-04-29T12:29:18.615Z + 2025-04-29T14:02:28.450Z https://docs.axolotl.ai/docs/api/prompt_strategies.kto.user_defined.html - 2025-04-29T12:29:18.731Z + 2025-04-29T14:02:28.567Z https://docs.axolotl.ai/docs/api/core.training_args.html - 2025-04-29T12:29:18.316Z + 2025-04-29T14:02:28.150Z https://docs.axolotl.ai/docs/api/utils.collators.mm_chat.html - 2025-04-29T12:29:19.349Z + 2025-04-29T14:02:29.190Z https://docs.axolotl.ai/docs/api/common.architectures.html - 2025-04-29T12:29:19.301Z + 2025-04-29T14:02:29.142Z https://docs.axolotl.ai/docs/api/integrations.kd.trainer.html - 2025-04-29T12:29:19.290Z + 2025-04-29T14:02:29.130Z https://docs.axolotl.ai/docs/api/prompt_strategies.llama2_chat.html - 2025-04-29T12:29:18.649Z + 2025-04-29T14:02:28.485Z https://docs.axolotl.ai/docs/api/core.trainers.trl.html - 2025-04-29T12:29:18.569Z + 2025-04-29T14:02:28.404Z https://docs.axolotl.ai/docs/api/prompt_strategies.orcamini.html - 2025-04-29T12:29:18.676Z + 2025-04-29T14:02:28.512Z https://docs.axolotl.ai/docs/api/integrations.cut_cross_entropy.args.html - 2025-04-29T12:29:19.281Z + 2025-04-29T14:02:29.122Z https://docs.axolotl.ai/docs/api/utils.lora.html - 2025-04-29T12:29:19.029Z + 2025-04-29T14:02:28.869Z https://docs.axolotl.ai/docs/api/cli.sweeps.html - 2025-04-29T12:29:18.493Z + 2025-04-29T14:02:28.327Z https://docs.axolotl.ai/docs/api/core.chat.format.chatml.html - 2025-04-29T12:29:18.340Z + 2025-04-29T14:02:28.174Z https://docs.axolotl.ai/docs/api/common.const.html - 2025-04-29T12:29:19.303Z + 2025-04-29T14:02:29.143Z https://docs.axolotl.ai/docs/api/prompt_strategies.metharme.html - 2025-04-29T12:29:18.673Z + 2025-04-29T14:02:28.508Z https://docs.axolotl.ai/docs/api/monkeypatch.stablelm_attn_hijack_flash.html - 2025-04-29T12:29:18.958Z + 2025-04-29T14:02:28.795Z https://docs.axolotl.ai/docs/api/integrations.grokfast.optimizer.html - 2025-04-29T12:29:19.282Z + 2025-04-29T14:02:29.123Z https://docs.axolotl.ai/docs/api/cli.checks.html - 2025-04-29T12:29:18.427Z + 2025-04-29T14:02:28.261Z https://docs.axolotl.ai/docs/api/common.datasets.html - 2025-04-29T12:29:19.319Z + 2025-04-29T14:02:29.160Z https://docs.axolotl.ai/docs/api/integrations.lm_eval.args.html - 2025-04-29T12:29:19.296Z + 2025-04-29T14:02:29.137Z https://docs.axolotl.ai/docs/api/monkeypatch.mixtral.html - 2025-04-29T12:29:18.979Z + 2025-04-29T14:02:28.818Z https://docs.axolotl.ai/docs/api/prompt_strategies.kto.chatml.html - 2025-04-29T12:29:18.730Z + 2025-04-29T14:02:28.565Z https://docs.axolotl.ai/docs/api/cli.train.html - 2025-04-29T12:29:18.396Z + 2025-04-29T14:02:28.229Z https://docs.axolotl.ai/docs/api/prompt_tokenizers.html - 2025-04-29T12:29:18.209Z + 2025-04-29T14:02:28.042Z https://docs.axolotl.ai/docs/api/utils.schemas.training.html - 2025-04-29T12:29:19.148Z + 2025-04-29T14:02:28.987Z https://docs.axolotl.ai/docs/api/monkeypatch.relora.html - 2025-04-29T12:29:18.916Z + 2025-04-29T14:02:28.752Z https://docs.axolotl.ai/docs/api/cli.cloud.modal_.html - 2025-04-29T12:29:18.538Z + 2025-04-29T14:02:28.373Z https://docs.axolotl.ai/docs/api/cli.main.html - 2025-04-29T12:29:18.387Z + 2025-04-29T14:02:28.221Z https://docs.axolotl.ai/docs/api/monkeypatch.trainer_fsdp_optim.html - 2025-04-29T12:29:18.962Z + 2025-04-29T14:02:28.798Z https://docs.axolotl.ai/docs/fsdp_qlora.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/debugging.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/ray-integration.html - 2025-04-29T12:28:44.173Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/docker.html - 2025-04-29T12:28:44.170Z + 2025-04-29T14:01:58.460Z https://docs.axolotl.ai/docs/inference.html - 2025-04-29T12:28:44.173Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/docs/torchao.html - 2025-04-29T12:28:44.174Z + 2025-04-29T14:01:58.463Z https://docs.axolotl.ai/FAQS.html - 2025-04-29T12:28:44.168Z + 2025-04-29T14:01:58.458Z