license: cc-by-nc-4.0

• 论文: RLHF工作流：从奖励建模到在线RLHF（发表于TMLR 2024）
• 作者: 董翰泽*、熊威*、庞博*、王浩翔*、韩昭、周英博、江南、Doyen Sahoo、熊才明、张潼
• 代码: https://github.com/RLHFlow/RLHF-Reward-Modeling/

该奖励函数可用于RLHF流程，包括PPO、迭代SFT和迭代DPO方法。

本许可证继承自PKU-Alignment/PKU-SafeRLHF-30K。

训练

基础模型采用meta-llama/Meta-Llama-3-8B-Instruct。

训练脚本见https://github.com/WeiXiongUST/RLHF-Reward-Modeling。

使用方法

  from transformers import AutoTokenizer, pipeline
  rm_tokenizer = AutoTokenizer.from_pretrained("sfairXC/FsfairX-LLaMA3-RM-v0.1")
  device = 0 # 加速器设备
  rm_pipe = pipeline(
      "sentiment-analysis",
      model="sfairXC/FsfairX-LLaMA3-RM-v0.1",
      #device="auto",
      device=device,
      tokenizer=rm_tokenizer,
      model_kwargs={"torch_dtype": torch.bfloat16}
  )

  pipe_kwargs = {
      "return_all_scores": True,
      "function_to_apply": "none",
      "batch_size": 1
  }

  chat = [
   {"role": "user", "content": "你好，最近怎么样？"},
   {"role": "assistant", "content": "我很好。今天能为您做些什么？"},
   {"role": "user", "content": "我想展示聊天模板的工作原理！"},
  ]

  test_texts = [rm_tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=False).replace(rm_tokenizer.bos_token, "")]
  pipe_outputs = rm_pipe(test_texts, **pipe_kwargs)
  rewards = [output[0]["score"] for output in pipe_outputs]

性能表现

本奖励模型是Reward-Bench榜单上当前（2024年4月20日）最先进的开源RM模型。

参考文献

本仓库是迭代拒绝采样微调与迭代DPO研究的一部分。若您的研究工作使用了本仓库内容，请参考以下引用格式：

@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}
}

@misc{xiong2024iterative,
      title={Iterative Preference Learning from Human Feedback: Bridging Theory and Practice for RLHF under KL-Constraint}, 
      author={Wei Xiong and Hanze Dong and Chenlu Ye and Ziqi Wang and Han Zhong and Heng Ji and Nan Jiang and Tong Zhang},
      year={2024},
      eprint={2312.11456},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}