Hugging Face's logo

berkeley-nest
/
Starling-LM-7B-alpha

Text Generation
Transformers
Safetensors
English
mistral
reward model
RLHF
RLAIF
conversational
text-generation-inference
Inference Endpoints
Model card Files Community
32
Starling-LM-7B-alpha / README.md
banghua's picture
banghua
Update README.md
d10bd8a
preview code
|
raw
history blame
5.56 kB
metadata
license: cc-by-nc-4.0
datasets:
  - berkeley-nest/Nectar
language:
  - en
library_name: transformers
tags:
  - reward model
  - RLHF
  - RLAIF

Starling-RM-7B-alpha

  • Developed by: Banghua Zhu * , Evan Frick * , Tianhao Wu * , Hanlin Zhu and Jiantao Jiao.
  • Model type: Language Model finetuned with RLHF / RLAIF
  • License: Non commercial license
  • Finetuned from model: Mistral-7B-v0.1

We introduce Starling-7B, an open large language model (LLM) trained by Reinforcement Learning from AI Feedback (RLAIF). The model harnesses the power of our new GPT-4 labeled ranking dataset, Nectar, and our new reward training and policy tuning pipeline. Starling-7B-alpha scores 8.09 in MT Bench with GPT-4 as a judge, outperforming every model to date on MT-Bench except for OpenAI's GPT-4 and GPT-4 Turbo. We release the ranking dataset Nectar, the reward model Starling-RM-7B-alpha and the language model Starling-LM-7B-alpha on HuggingFace, and an online demo in LMSYS Chatbot Arena. Stay tuned for our forthcoming code and paper, which will provide more details on the whole process.

Starling-LM-7B-alpha is a language model trained from Openchat 3.5

Model Tuning Method MT Bench AlpacaEval MMLU
GPT-4-Turbo ? 9.32 97.70
GPT-4 SFT + PPO 8.99 95.28 86.4
Starling-7B C-RLFT + APA 8.09 91.99 63.9
Claude-2 ? 8.06 91.36 78.5
GPT-3.5-Turbo ? 7.94 89.37 70
Claude-1 ? 7.9 88.39 77
Tulu-2-dpo-70b SFT + DPO 7.89 95.1
Openchat-3.5 C-RLFT 7.81 88.51 64.3
Zephyr-7B-beta SFT + DPO 7.34 90.60 61.4
Llama-2-70b-chat-hf SFT + PPO 6.86 92.66 63
Neural-chat-7b-v3-1 SFT + DPO 6.84 84.53 62.4
Tulu-2-dpo-7b SFT + DPO 6.29 85.1

ollowing the method of training reward model in the instructGPT paper, we remove the last layer of Llama2-7B Chat, and concatenate a linear layer that outputs scalar for any pair of input prompt and response. We train the reward model with preference dataset berkeley-nest/Nectar, with the K-wise maximum likelihood estimator proposed in this paper. The reward model outputs a scalar for any given prompt and response. A response that is more helpful and less harmful will get the highest reward score. Note that since the preference dataset berkeley-nest/Nectar is based on GPT-4 preference, the reward model is likely to be biased towards GPT-4's own preference, including longer responses and certain response format.

For more detailed discussions, please check out our blog post, and stay tuned for our upcoming code and paper!

Uses

Our model follows the exact chat template and usage as Openchat 3.5. Please refer to their model card for more details. In addition, our model is hosted on LMSYS Chatbot Arena for free test.

License

The dataset, model and online demo is a research preview intended for non-commercial use only, subject to the data distillation License of LLaMA, Terms of Use of the data generated by OpenAI, and Privacy Practices of ShareGPT. Please contact us if you find any potential violation.

Acknowledgment

We would like to thank Wei-Lin Chiang from Berkeley for detailed feedback of the blog and the projects. We would like to thank the LMSYS Organization for their support of lmsys-chat-1M dataset, evaluation and online demo. We would like to thank the open source community for their efforts in providing the datasets and base models we used to develope the project, including but not limited to Anthropic, Llama, Mistral, Hugging Face H4, LMSYS, OpenChat, OpenBMB, Flan and ShareGPT.

Citation 

@misc{starling2023,
    title = {Starling-7B: Improving LLM Helpfulness & Harmlessness with RLAIF},
    url = {},
    author = {Zhu, Banghua and Frick, Evan and Wu, Tianhao and Zhu, Hanlin and Jiao, Jiantao},
    month = {November},
    year = {2023}
}