116 lines
5.8 KiB
Markdown
116 lines
5.8 KiB
Markdown
---
|
|
license: apache-2.0
|
|
library_name: transformers
|
|
tags:
|
|
- storm
|
|
- mistral
|
|
- openchat
|
|
- RLAIF
|
|
- reward model
|
|
language:
|
|
- en
|
|
base_model: openchat/openchat-3.5-0106
|
|
datasets:
|
|
- berkeley-nest/Nectar
|
|
---
|
|
|
|
# Storm-7B
|
|
- **Developed by**: [Jie Liu](https://jieliu.site/) \\(^{*1,2}\\), [Zhanhui Zhou](https://scholar.google.com/citations?user=SbACfYQAAAAJ&hl=zh-CN) \\(^{*2}\\), [Jiaheng Liu](https://liujiaheng.github.io/) \\(^{2}\\), [Xingyuan Bu](https://scholar.google.com.hk/citations?user=cqYaRhUAAAAJ&hl=zh-CN) \\(^{2}\\), [Chao Yang](https://scholar.google.com/citations?user=5KRbHPMAAAAJ&hl=zh-CN) \\(^{2}\\), [Han-Sen Zhong](https://scholar.google.com.hk/citations?user=X_ZfX8sAAAAJ&hl=zh-CN) \\(^{\dag 2}\\), [Wanli Ouyang](https://wlouyang.github.io/) \\(^{1,2}\\).
|
|
- \\(^{1}\\)MMLab, The Chinese University of Hong Kong   \\(^{2}\\)Shanghai AI Laboratory
|
|
- Paper: [Iterative Length-Regularized Direct Preference Optimization: A Case Study on Improving 7B Language Models to GPT-4 Level](https://arxiv.org/pdf/2406.11817)
|
|
- Finetuned from the model: [openchat-3.5-0106](https://huggingface.co/openchat/openchat-3.5-0106)
|
|
- Dataset: [berkeley-nest/Nectar](https://huggingface.co/datasets/berkeley-nest/Nectar)
|
|
- Reward Model: [Starling-RM-34B](https://huggingface.co/Nexusflow/Starling-RM-34B)
|
|
|
|
Please see our paper for more details.
|
|
|
|
## Introduction
|
|
|
|
We released Storm-7B, the first open-source language model comparable to the GPT-4 series on the [AlpacaEval 2.0](https://tatsu-lab.github.io/alpaca_eval/) leaderboard.
|
|
|
|
Recent studies show that DPO benefits from iterative training with online preferences labeled by a trained reward model. In this work, we identify a pitfall of vanilla iterative DPO - improved response quality can lead to increased verbosity. To address this, we introduce iterative length-regularized DPO (iLR-DPO) to penalize response length. Our empirical results show that iLR-DPO can enhance a 7B model to perform on par with GPT-4 **without increasing verbosity**.
|
|
|
|
## Performance
|
|
Our 7B model achieves a **50.5%** length-controlled win rate against GPT-4 Preview on AlpacaEval 2.0.
|
|
<p align="center">
|
|
<img src="https://cdn-uploads.huggingface.co/production/uploads/639be86b59473c6ae02ef9c4/Tj_a1QntAxkhy2SXbOdmT.png" width="60%">
|
|
</p>
|
|
Our model's LC win rate improves over iterations without significantly changing the response length, indicating better alignment with human values without length bias. The final trained model (iteration 3) achieves a 50.5% LC win rate, making it the first open-source model to surpass the baseline model GPT-4 Preview.
|
|
|
|
In addition to regular decoding, we also test beam search and best-of-n sampling on top of our trained model. Beam search over our trained model shows a 5% improvement over regular decoding, Best-of-n sampling with Starling-RM-34B achieves 61.6% LC Win rate and outperforms GPT-4 Omni.
|
|
<p align="center">
|
|
<img src="https://cdn-uploads.huggingface.co/production/uploads/639be86b59473c6ae02ef9c4/GGa28vaREaVq099MPdqcP.png" width="100%">
|
|
</p>
|
|
|
|
We observe no significant degradation in traditional NLP tasks from the Huggingface Open LLM Leaderboard.
|
|
<p align="center">
|
|
<img src="https://cdn-uploads.huggingface.co/production/uploads/639be86b59473c6ae02ef9c4/8KEm_Ladg7Kqko8mC63SN.png" width="100%">
|
|
</p>
|
|
|
|
|
|
## Uses
|
|
|
|
Our model uses the same chat template as [Openchat-3.5-0106](https://huggingface.co/openchat/openchat-3.5-0106). A sample code snippet for inference using our model is provided below.
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
device = "cuda"
|
|
|
|
model = AutoModelForCausalLM.from_pretrained("jieliu/Storm-7B").to(device)
|
|
tokenizer = AutoTokenizer.from_pretrained("jieliu/Storm-7B")
|
|
model.eval().requires_grad_(False)
|
|
|
|
def generate_response(prompt):
|
|
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
|
|
outputs = model.generate(
|
|
input_ids,
|
|
max_length=2048,
|
|
do_sample=True,
|
|
temperature=1.0,
|
|
pad_token_id=tokenizer.pad_token_id,
|
|
eos_token_id=tokenizer.eos_token_id,
|
|
)
|
|
response_ids = outputs[0]
|
|
response_text = tokenizer.decode(response_ids, skip_special_tokens=True)
|
|
return response_text
|
|
|
|
prompt = "How does a telescope work?"
|
|
input_prompt = f"GPT4 Correct User: {prompt}<|end_of_turn|>GPT4 Correct Assistant:"
|
|
response_text = generate_response(input_prompt)
|
|
print("Response:", response_text)
|
|
```
|
|
|
|
## Scripts
|
|
You can reproduce our results on AlphaEval 2.0 using the script provided below.
|
|
```bash
|
|
git clone https://github.com/tatsu-lab/alpaca_eval.git
|
|
cd alpaca_eval
|
|
pip install -e .
|
|
export OPENAI_API_KEY=<your_api_key>
|
|
alpaca_eval evaluate_from_model --model_configs 'Storm-7B'
|
|
```
|
|
|
|
## Limitations
|
|
|
|
Our work has several limitations:
|
|
(1) We focus on aligning with human preferences but only use GPT-4 as a proxy for human judgment to evaluate language models.
|
|
(2) We reduce verbosity with a length penalty, though verbosity and length are not necessarily correlated. Future work could train a specific reward model to directly penalize verbosity, replacing the length margin with a verbosity margin, following the standard [MODPO pipeline](https://github.com/ZHZisZZ/modpo).
|
|
|
|
## Citation
|
|
|
|
```
|
|
@article{liu2024iterative,
|
|
title = {Iterative Length-Regularized Direct Preference Optimization: A Case Study on Improving 7B Language Models to GPT-4 Level},
|
|
author = {Liu, Jie and Zhou, Zhanhui and Liu, Jiaheng and Bu, Xingyuan and Yang, Chao and Zhong Han-Sen and Ouyang, Wanli},
|
|
journal={arXiv preprint arXiv:2406.11817},
|
|
year={2024}
|
|
}
|
|
|
|
@article{zhou2023beyond,
|
|
title={Beyond one-preference-for-all: Multi-objective direct preference optimization},
|
|
author={Zhou, Zhanhui and Liu, Jie and Yang, Chao and Shao, Jing and Liu, Yu and Yue, Xiangyu and Ouyang, Wanli and Qiao, Yu},
|
|
journal={arXiv preprint arXiv:2310.03708},
|
|
year={2023}
|
|
}
|
|
``` |