base_model: openai/gpt-oss-20b
use_kernels: true
model_quantization_config: Mxfp4Config
model_quantization_config_kwargs:
  dequantize: true

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

experimental_skip_move_to_device: true  # prevent OOM by not putting model to GPU before sharding

datasets:
  - path: HuggingFaceH4/Multilingual-Thinking
    type: chat_template
    field_thinking: thinking
    template_thinking_key: thinking

dataset_prepared_path: last_run_prepared
val_set_size: 0
output_dir: ./outputs/gpt-oss-out/

sequence_len: 4096
sample_packing: true

adapter: lora
lora_r: 8
lora_alpha: 16
lora_dropout: 0.0  # dropout not supported when using LoRA over expert parameters
lora_target_linear: true

# TODO: not supported for now, see peft#2710
#lora_target_parameters:  # target the experts in the last two layers
#  - "22._checkpoint_wrapped_module.mlp.experts.gate_up_proj"
#  - "22._checkpoint_wrapped_module.mlp.experts.down_proj"
#  - "23._checkpoint_wrapped_module.mlp.experts.gate_up_proj"
#  - "23._checkpoint_wrapped_module.mlp.experts.down_proj"

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

trackio_project_name:
trackio_run_name:
trackio_space_id:

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 1

optimizer: adamw_torch_8bit
lr_scheduler: constant_with_warmup
learning_rate: 2e-4

bf16: true
tf32: true

flash_attention: true
attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3

gradient_checkpointing: true
activation_offloading: true

logging_steps: 1
saves_per_epoch: 1
warmup_ratio: 0.1

special_tokens:
eot_tokens:
  - "<|end|>"