| # Model arguments | |
| model_name_or_path: mistralai/Mistral-7B-v0.1 | |
| model_revision: main | |
| torch_dtype: bfloat16 | |
| use_flash_attention_2: true | |
| # Data training arguments | |
| dataset_mixer: | |
| HuggingFaceH4/ultrachat_200k: 1.0 | |
| dataset_splits: | |
| - train_sft | |
| - test_sft | |
| preprocessing_num_workers: 12 | |
| # SFT trainer config | |
| bf16: true | |
| do_eval: true | |
| evaluation_strategy: epoch | |
| gradient_accumulation_steps: 2 | |
| gradient_checkpointing: true | |
| hub_model_id: zephyr-7b-sft-full | |
| hub_strategy: every_save | |
| learning_rate: 2.0e-05 | |
| log_level: info | |
| logging_steps: 5 | |
| logging_strategy: steps | |
| lr_scheduler_type: cosine | |
| max_seq_length: 2048 | |
| max_steps: -1 | |
| num_train_epochs: 1 | |
| output_dir: data/zephyr-7b-sft-full | |
| overwrite_output_dir: true | |
| per_device_eval_batch_size: 16 | |
| per_device_train_batch_size: 32 | |
| push_to_hub: true | |
| remove_unused_columns: true | |
| report_to: | |
| - tensorboard | |
| save_strategy: "no" | |
| save_total_limit: null | |
| seed: 42 |