QuantFactory/LLaMA-3-8B-SFR-Iterative-DPO-R-GGUF
1568
1---2license: llama33pipeline_tag: text-generation4base_model: Salesforce/LLaMA-3-8B-SFR-Iterative-DPO-R5---6# Llama-3-8B-SFR-Iterative-DPO-R-GGUF7This is quantized version of [Salesforce/LLaMA-3-8B-SFR-Iterative-DPO-R](https://huggingface.co/Salesforce/LLaMA-3-8B-SFR-Iterative-DPO-R) created using llama.cpp8 9## Model Description10We release a state-of-the-art instruct model of its class, **Llama-3-8B-SFR-Iterative-DPO-R**.11On 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),12and strong proprietary models (e.g., GPT-3.5-turbo-0613). The model is trained with open-sourced datasets without any additional human-/GPT4-labeling.13 14## Model Releases15- [SFT model](https://huggingface.co/Salesforce/SFR-SFT-LLaMA-3-8B-R)16- [Reward model](https://huggingface.co/Salesforce/SFR-RM-LLaMA-3-8B-R)17- [RLHF model](https://huggingface.co/Salesforce/SFR-Iterative-DPO-LLaMA-3-8B-R)18 19 20## Training methods21We 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.22Unlike widely-used offline DPO, the online component of our approach effectively mitigates distribution shifts during policy optimization.23For a detailed exposition, please refer to our accompanying technical report.24 25 26## Chat Benchmarks27 28| **Model** | **Size** | **Method** | **LC Alpaca-Eval-V2** | **MT-Bench** | **Chat-Arena-Hard** |29|-------------------------|----------|-------------------|-----------------------|--------------|---------------------|30| **Small Open-Sourced Models** | | | | | |31| Gemma-7B-it | 7B | SFT | 10.4 | 6.38 | 7.5 |32| Zephyr-7B-beta | 7B | Vanilla DPO | 13.1 | 7.34 | - |33| Mistral-7B-v0.2-it | 7B | SFT | 17.1 | 7.51 | 12.6 |34| Open-Chat-0106 | 7B | SFT | 15.6 | 7.8 | - |35| Starling-7B-beta | 7B | PPO | 25.8 | 8.12 | 23.0 |36| LLaMA-3-8B-it | 8B | RS+DPO+PPO | 22.9 | 8.16 | 20.6 |37| **Ours** | | | | | |38| Ours (SFT baseline) | 8B | SFT | 10.2 | 7.69 | 5.6 |39| Ours (DPO baseline) | 8B | Vanilla DPO | 22.5 | 8.17 | 22.4 |40| Ours (Online RLHF) | 8B | Iterative DPO | **31.3** | **8.46** | **29.1** |41| **Large Open-Sourced Models** | | | | | |42| Vicuna-33b-v1.3 | 33B | SFT | 17.6 | 7.12 | 8.6 |43| Yi-34B-Chat | 34B | SFT | 27.2 | - | 23.1 |44| Mixtral-8x7B-it | 45B* | SFT | 23.7 | 8.30 | 23.4 |45| Tulu-2-DPO-70B | 70B | Vanilla DPO | 21.2 | 7.89 | 15.0 |46| LLaMA-3-70B-it | 70B | RS+DPO+PPO | 34.4 | 8.95 | 41.1 |47| Mixtral-8x22B-it | 141B* | SFT | 30.9 | 8.66 | 36.4 |48| **Proprietary Models** | | | | | |49| GPT-3.5-turbo-1106 | - | - | 19.3 | 8.35 | 18.9 |50| GPT-3.5-turbo-0613 | - | - | 22.7 | 8.39 | 24.8 |51| GPT-4-0613 | - | - | 30.2 | 9.18 | 37.9 |52| Claude-3-Opus | - | - | 40.5 | 9.00 | 60.4 |53| GPT-4 Turbo (04/09) | - | - | 55.0 | - | 82.6 |54 55 56## Academic Benchmarks57 58| **Model** | **Size** | **Method** | **GSM-8K** | **MMLU** | **HumanEval** | **TruthfulQA** | **ARC** | **MBPP** |59|----------------------------|----------|-----------------|------------|----------|---------------|----------------|---------|----------|60| LLaMA-3-8B-it | 8B | RS+DPO+PPO | 79.6 | 66.0 | 61.6 | 43.9 | 59.5 | 61.1 |61| Ours (SFT baseline) | 8B | SFT | 74.2 | 64.7 | 65.2 | 53.4 | 61.4 | 62.3 |62| Ours (DPO baseline) | 8B | Vanilla DPO | 79.8 | 64.5 | 63.4 | 61.8 | 65.2 | 60.3 |63| Ours (Iterative RLHF) | 8B | Iterative DPO | 80.7 | 65.3 | 64.6 | 60.4 | 64.3 | 60.8 |64 65 66## Usage67```python68from transformers import AutoModelForCausalLM, AutoTokenizer69 70device = "cuda" 71 72model = AutoModelForCausalLM.from_pretrained("Salesforce/Llama-3-8B-SFR-Iterative-DPO-R")73tokenizer = AutoTokenizer.from_pretrained("Salesforce/Llama-3-8B-SFR-Iterative-DPO-R")74 75messages = [76 {"role": "user", "content": "I'm trying to teach myself to have nicer handwriting. Can you help?"},77]78 79model_inputs = tokenizer.apply_chat_template(messages, return_tensors="pt")80 81model_inputs = model_inputs.to(device)82model.to(device)83 84output_tokens = model.generate(model_inputs, max_new_tokens=1024, do_sample=True)85model_outputs = tokenizer.batch_decode(output_tokens)86print(model_outputs[0])87```88 89 90## Limitations91Llama-3-8B-SFR-Iterative-DPO-R is a research model developed as part of our RLHF initiative at Salesforce. 92While safety and ethical considerations are integral to our alignment process, 93there remains the possibility that the model could generate offensive or unethical content, particularly under adversarial conditions. 94We are committed to continuous improvement in our models to minimize such risks and encourage responsible usage.95 96## Original Model Citation97Please cite our papers if you find our models are useful.98 99```bibtex100@misc{dong2024rlhf,101 title={RLHF Workflow: From Reward Modeling to Online RLHF}, 102 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},103 year={2024},104 eprint={2405.07863},105 archivePrefix={arXiv},106 primaryClass={cs.LG}107}108 109@misc{xiong2024iterative,110 title={Iterative Preference Learning from Human Feedback: Bridging Theory and Practice for RLHF under KL-Constraint}, 111 author={Wei Xiong and Hanze Dong and Chenlu Ye and Ziqi Wang and Han Zhong and Heng Ji and Nan Jiang and Tong Zhang},112 year={2024},113 eprint={2312.11456},114 archivePrefix={arXiv},115 primaryClass={cs.LG}116}117```