RLHFlow/LLaMA3-iterative-DPO-final
45208
1---2license: llama33---4# LLaMA3-iterative-DPO-final5 6* **Paper**: [RLHF Workflow: From Reward Modeling to Online RLHF](https://arxiv.org/pdf/2405.07863) (Published in TMLR, 2024)7* **Authors**: Hanze Dong*, Wei Xiong*, Bo Pang*, Haoxiang Wang*, Han Zhao, Yingbo Zhou, Nan Jiang, Doyen Sahoo, Caiming Xiong, Tong Zhang8* **Code**: https://github.com/RLHFlow/Online-RLHF 9 10## Introduction11We release an unofficial checkpoint of a state-of-the-art instruct model of its class, **LLaMA3-iterative-DPO-final**.12On all three widely-used instruct model benchmarks: **Alpaca-Eval-V2**, **MT-Bench**, **Chat-Arena-Hard**, our model outperforms all models of similar size (e.g., LLaMA-3-8B-it), most large open-sourced models (e.g., Mixtral-8x7B-it),13and strong proprietary models (e.g., GPT-3.5-turbo-0613). The model is trained with open-sourced datasets without any additional human-/GPT4-labeling.14 15Even better, we provide a [detailed recipe](https://github.com/RLHFlow/Online-RLHF) to reproduce the model. Enjoy!16 17## Model Releases18See the [collection](https://huggingface.co/collections/RLHFlow/online-rlhf-663ae95fade1a39663dab218) of the training set, reward/preference model, SFT model.19 20- [SFT model](https://huggingface.co/RLHFlow/LLaMA3-SFT)21- [Reward model](https://huggingface.co/sfairXC/FsfairX-LLaMA3-RM-v0.1)22- This model is more like the concise version in the report. We are still working on the model realeasing due to some license issue....23 24## Dataset 25- [Preference data mix](https://huggingface.co/datasets/hendrydong/preference_700K)26- [Prompt collection for RLHF training](https://huggingface.co/datasets/RLHFlow/prompt-collection-v0.1)27 28## Training methods29We have developed a simple and efficient online RLHF recipe for LLM instruct training. Our recipe is DPO-based and thus much cheaper and simpler to train and tune compared to PPO-based approaches.30Unlike widely-used offline DPO, the online component of our approach effectively mitigates distribution shifts during policy optimization.31For a detailed exposition, please refer to our accompanying technical report.32 33 34## Chat Benchmarks35 36| **Model** | **Size** | **Method** | **LC Alpaca-Eval-V2** | **MT-Bench** | **Chat-Arena-Hard** |37|-------------------------|----------|-------------------|-----------------------|--------------|---------------------|38| **Small Open-Sourced Models** | | | | | |39| Gemma-7B-it | 7B | SFT | 10.4 | 6.38 | 7.5 |40| Zephyr-7B-beta | 7B | Vanilla DPO | 13.1 | 7.34 | - |41| Mistral-7B-v0.2-it | 7B | SFT | 17.1 | 7.51 | 12.6 |42| Open-Chat-0106 | 7B | SFT | 15.6 | 7.8 | - |43| Starling-7B-beta | 7B | PPO | 25.8 | 8.12 | 23.0 |44| LLaMA-3-8B-it | 8B | RS+DPO+PPO | 22.9 | 8.16 | 20.6 |45| **Ours** | | | | | |46| Ours (SFT baseline) | 8B | SFT | 10.2 | 7.69 | 5.6 |47| Ours (DPO baseline) | 8B | Vanilla DPO | 22.5 | 8.17 | 22.4 |48| Ours (Online RLHF) | 8B | Iterative DPO | **37.2** | **8.46** | **29.1** |49| **Large Open-Sourced Models** | | | | | |50| Vicuna-33b-v1.3 | 33B | SFT | 17.6 | 7.12 | 8.6 |51| Yi-34B-Chat | 34B | SFT | 27.2 | - | 23.1 |52| Mixtral-8x7B-it | 45B* | SFT | 23.7 | 8.30 | 23.4 |53| Tulu-2-DPO-70B | 70B | Vanilla DPO | 21.2 | 7.89 | 15.0 |54| LLaMA-3-70B-it | 70B | RS+DPO+PPO | 34.4 | 8.95 | 41.1 |55| Mixtral-8x22B-it | 141B* | SFT | 30.9 | 8.66 | 36.4 |56| **Proprietary Models** | | | | | |57| GPT-3.5-turbo-1106 | - | - | 19.3 | 8.35 | 18.9 |58| GPT-3.5-turbo-0613 | - | - | 22.7 | 8.39 | 24.8 |59| GPT-4-0613 | - | - | 30.2 | 9.18 | 37.9 |60| Claude-3-Opus | - | - | 40.5 | 9.00 | 60.4 |61| GPT-4 Turbo (04/09) | - | - | 55.0 | - | 82.6 |62 63 64## Academic Benchmarks65 66| **Model** | **Size** | **Method** | **GSM-8K** | **MMLU** | **HumanEval** | **TruthfulQA** | **ARC** | **MBPP** |67|----------------------------|----------|-----------------|------------|----------|---------------|----------------|---------|----------|68| LLaMA-3-8B-it | 8B | RS+DPO+PPO | 79.6 | 66.0 | 61.6 | 43.9 | 59.5 | 61.1 |69| Ours (SFT baseline) | 8B | SFT | 74.2 | 64.7 | 65.2 | 53.4 | 61.4 | 62.3 |70| Ours (DPO baseline) | 8B | Vanilla DPO | 79.8 | 64.5 | 63.4 | 61.8 | 65.2 | 60.3 |71| Ours (Iterative RLHF) | 8B | Iterative DPO | 80.7 | 65.3 | 64.6 | 60.4 | 64.3 | 60.8 |72 73 74## Usage75```python76from transformers import AutoModelForCausalLM, AutoTokenizer77 78device = "cuda" 79 80model = AutoModelForCausalLM.from_pretrained("RLHFlow/LLaMA3-iterative-DPO-final")81tokenizer = AutoTokenizer.from_pretrained("RLHFlow/LLaMA3-iterative-DPO-final")82 83messages = [84 {"role": "user", "content": "I'm trying to teach myself to have nicer handwriting. Can you help?"},85]86 87model_inputs = tokenizer.apply_chat_template(messages, return_tensors="pt")88 89model_inputs = model_inputs.to(device)90model.to(device)91 92output_tokens = model.generate(model_inputs, max_new_tokens=1024, do_sample=True)93model_outputs = tokenizer.batch_decode(output_tokens)94print(model_outputs[0])95```96 97 98## Limitations99RLHFlow/LLaMA3-iterative-DPO-final is an unofficial checkpoint developed to illustrate the power of online iterative RLHF and is for research purpose. While safety and ethical considerations are integral to our alignment process, 100there remains the possibility that the model could generate offensive or unethical content, particularly under adversarial conditions. 101We are committed to continuous improvement in our models to minimize such risks and encourage responsible usage.102 103## Citation104Please cite our techical report if you find our model is useful for your research or product.105```106@misc{dong2024rlhf,107 title={RLHF Workflow: From Reward Modeling to Online RLHF}, 108 author={Hanze Dong and Wei Xiong and Bo Pang and Haoxiang Wang and Han Zhao and Yingbo Zhou and Nan Jiang and Doyen Sahoo and Caiming Xiong and Tong Zhang},109 year={2024},110 eprint={2405.07863},111 archivePrefix={arXiv},112 primaryClass={cs.LG}113}114 115@misc{xiong2024iterative,116 title={Iterative Preference Learning from Human Feedback: Bridging Theory and Practice for RLHF under KL-Constraint}, 117 author={Wei Xiong and Hanze Dong and Chenlu Ye and Ziqi Wang and Han Zhong and Heng Ji and Nan Jiang and Tong Zhang},118 year={2024},119 eprint={2312.11456},120 archivePrefix={arXiv},121 primaryClass={cs.LG}122}123 124```