| # Model arguments | |
| model_name_or_path: sanchit-gandhi/Mistral-7B-v0.1-6-layer | |
| model_revision: main | |
| torch_dtype: bfloat16 | |
| use_flash_attention_2: false # torch sdpa sufficient | |
| # Data training arguments | |
| chat_template: "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ '<|user|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'system' %}\n{{ '<|system|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'assistant' %}\n{{ '<|assistant|>\n' + message['content'] + eos_token }}\n{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ '<|assistant|>' }}\n{% endif %}\n{% endfor %}" | |
| dataset_mixer: | |
| stingning/ultrachat: 1.0 | |
| dataset_splits: | |
| - train[1000:] | |
| - train[:1000] | |
| preprocessing_num_workers: 32 | |
| # SFT trainer config | |
| bf16: true | |
| do_eval: true | |
| evaluation_strategy: steps | |
| eval_steps: 5000 | |
| save_strategy: "steps" | |
| save_steps: 5000 | |
| save_total_limit: 1 | |
| gradient_accumulation_steps: 1 | |
| gradient_checkpointing: true | |
| gradient_checkpointing_kwargs: | |
| use_reentrant: False | |
| hub_strategy: every_save | |
| learning_rate: 0.0001 | |
| log_level: info | |
| logging_steps: 25 | |
| logging_strategy: steps | |
| max_seq_length: 2048 | |
| max_steps: 20000 | |
| output_dir: ./ | |
| overwrite_output_dir: true | |
| per_device_eval_batch_size: 32 | |
| per_device_train_batch_size: 32 | |
| push_to_hub: true | |
| remove_unused_columns: true | |
| report_to: | |
| - tensorboard | |
| - wandb | |
| seed: 42 | |
| warmup_steps: 500 | |
| ddp_timeout: 7200 | |