Storm-7B
- Developed by: Jie Liu \(^{*1,2}\), Zhanhui Zhou \(^{*2}\), Jiaheng Liu \(^{2}\), Xingyuan Bu \(^{2}\), Chao Yang \(^{2}\), Han-Sen Zhong \(^{\dag 2}\), Wanli Ouyang \(^{1,2}\).
- \(^{1}\)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
- Finetuned from the model: openchat-3.5-0106
- Dataset: berkeley-nest/Nectar
- Reward Model: 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 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.
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.
We observe no significant degradation in traditional NLP tasks from the Huggingface Open LLM Leaderboard.
Uses
Our model uses the same chat template as Openchat-3.5-0106. A sample code snippet for inference using our model is provided below.
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.
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.
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}
}