HF_Trainer is a user-friendly and extensible framework for large-model training based on Huggingface. It supports various mainstream open-source large model training tasks, including pre-training, fine-tuning, reward modeling, and DPO training.
You can find all necessary parameter settings under the /config directory, including data_args.sh for dataset construction, train_args.sh for training, and eval_args.sh for evaluation.
To identify parameters that need to be set, search for YOUR_ in the configuration files. These parameters include DATASET_PATH, MODEL_PATH, SAVE_PATH, and WANDB_KEY (if not using W&B, you can disable it in the settings).
If you need to switch between different training stages, you can modify the training_stage parameter in train_args.sh.
When loading datasets, you can configure the corresponding fields by setting utils_config/data_column.json.
We provide two training scripts: scripts/train_local.sh for local single-machine training and scripts/train_submit.sh for distributed multi-machine training (using PyTorch DDP).
You can start training by running the command: bash scripts/train_xxxxx.sh.
Note: To save storage space, the framework automatically deletes optimizer parameters except for the latest checkpoint during training. If you want to save optimizer parameters for all checkpoints, comment out the line python tools/del_training_state.py ${STORE_DIR} $$ & in the script.
Similar to training, we provide two evaluation scripts for both local and distributed settings. You can start the evaluation by running the command: bash scripts/eval_xxxxx.sh.