Llama-3-8B-SFR-Iterative-DPO-R

Introduction

We release a state-of-the-art instruct model of its class, Llama-3-8B-SFR-Iterative-DPO-R. On 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), and strong proprietary models (e.g., GPT-3.5-turbo-0613). The model is trained with open-sourced datasets without any additional human-/GPT4-labeling.

Model Releases

• SFT model
• Reward model
• RLHF model

Training methods

We 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. Unlike widely-used offline DPO, the online component of our approach effectively mitigates distribution shifts during policy optimization. For a detailed exposition, please refer to our accompanying technical report.

Chat Benchmarks

Model	Size	Method	LC Alpaca-Eval-V2	MT-Bench	Chat-Arena-Hard
Small Open-Sourced Models
Gemma-7B-it	7B	SFT	10.4	6.38	7.5
Zephyr-7B-beta	7B	Vanilla DPO	13.1	7.34	-
Mistral-7B-v0.2-it	7B	SFT	17.1	7.51	12.6
Open-Chat-0106	7B	SFT	15.6	7.8	-
Starling-7B-beta	7B	PPO	25.8	8.12	23.0
LLaMA-3-8B-it	8B	RS+DPO+PPO	22.9	8.16	20.6
Ours
Ours (SFT baseline)	8B	SFT	10.2	7.69	5.6
Ours (DPO baseline)	8B	Vanilla DPO	22.5	8.17	22.4
Ours (Online RLHF)	8B	Iterative DPO	31.3	8.46	29.1
Large Open-Sourced Models
Vicuna-33b-v1.3	33B	SFT	17.6	7.12	8.6
Yi-34B-Chat	34B	SFT	27.2	-	23.1
Mixtral-8x7B-it	45B*	SFT	23.7	8.30	23.4
Tulu-2-DPO-70B	70B	Vanilla DPO	21.2	7.89	15.0
LLaMA-3-70B-it	70B	RS+DPO+PPO	34.4	8.95	41.1
Mixtral-8x22B-it	141B*	SFT	30.9	8.66	36.4
Proprietary Models
GPT-3.5-turbo-1106	-	-	19.3	8.35	18.9
GPT-3.5-turbo-0613	-	-	22.7	8.39	24.8
GPT-4-0613	-	-	30.2	9.18	37.9
Claude-3-Opus	-	-	40.5	9.00	60.4
GPT-4 Turbo (04/09)	-	-	55.0	-	82.6

Academic Benchmarks

Model	Size	Method	GSM-8K	MMLU	HumanEval	TruthfulQA	ARC	MBPP
LLaMA-3-8B-it	8B	RS+DPO+PPO	79.6	66.0	61.6	43.9	59.5	61.1
Ours (SFT baseline)	8B	SFT	74.2	64.7	65.2	53.4	61.4	62.3
Ours (DPO baseline)	8B	Vanilla DPO	79.8	64.5	63.4	61.8	65.2	60.3
Ours (Iterative RLHF)	8B	Iterative DPO	80.7	65.3	64.6	60.4	64.3	60.8

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

device = "cuda" 

model = AutoModelForCausalLM.from_pretrained("Salesforce/Llama-3-8B-SFR-Iterative-DPO-R")
tokenizer = AutoTokenizer.from_pretrained("Salesforce/Llama-3-8B-SFR-Iterative-DPO-R")

messages = [
    {"role": "user", "content": "I'm trying to teach myself to have nicer handwriting. Can you help?"},
]

model_inputs = tokenizer.apply_chat_template(messages, return_tensors="pt")

model_inputs = model_inputs.to(device)
model.to(device)

output_tokens = model.generate(model_inputs, max_new_tokens=1024, do_sample=True)
model_outputs = tokenizer.batch_decode(output_tokens)
print(model_outputs[0])

Limitations

Llama-3-8B-SFR-Iterative-DPO-R is a research model developed as part of our RLHF initiative at Salesforce. While safety and ethical considerations are integral to our alignment process, there remains the possibility that the model could generate offensive or unethical content, particularly under adversarial conditions. We are committed to continuous improvement in our models to minimize such risks and encourage responsible usage.

Citation

Please cite our papers if you find our models are useful.

@misc{dong2024rlhf,
      title={RLHF Workflow: From Reward Modeling to Online RLHF}, 
      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},
      year={2024},
      eprint={2405.07863},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

@misc{xiong2024iterative,
      title={Iterative Preference Learning from Human Feedback: Bridging Theory and Practice for RLHF under KL-Constraint}, 
      author={Wei Xiong and Hanze Dong and Chenlu Ye and Ziqi Wang and Han Zhong and Heng Ji and Nan Jiang and Tong Zhang},
      year={2024},
      eprint={2312.11456},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}