Training

The base model is mistralai/Mistral-7B-Instruct-v0.2.

We also merge the training script at https://github.com/WeiXiongUST/RLHF-Reward-Modeling.

Thanks Wei (https://huggingface.co/weqweasdas) for his help and contribution to the community.

Usage

To use this model, you need to load by AutoModelForSequenceClassification,

model = AutoModelForSequenceClassification.from_pretrained(
    "hendrydong/Mistral-RM-for-RAFT-GSHF-v0", num_labels=1, torch_dtype=torch.bfloat16
)

and prepare dataset like

SAMPLE =[
{'role': 'user', 'content': 'Hi!'},
{'role': 'assistant', 'content': 'How are you?'},
]

The template is the same as mistralai/Mistral-7B-Instruct-v0.2.

The reward model can be used for iterative SFT/DPO.

Please cite them if you found this RM helpful,

@article{dong2023raft,
  title={Raft: Reward ranked finetuning for generative foundation model alignment},
  author={Dong, Hanze and Xiong, Wei and Goyal, Deepanshu and Pan, Rui and Diao, Shizhe and Zhang, Jipeng and Shum, Kashun and Zhang, Tong},
  journal={arXiv preprint arXiv:2304.06767},
  year={2023}
}

@article{xiong2023gibbs,
  title={Gibbs sampling from human feedback: A provable kl-constrained framework for rlhf},
  author={Xiong, Wei and Dong, Hanze and Ye, Chenlu and Zhong, Han and Jiang, Nan and Zhang, Tong},
  journal={arXiv preprint arXiv:2312.11456},
  year={2023}
}