LLaMA3-iterative-DPO-final-ExPO

The extrapolated (ExPO) model based on RLHFlow/LLaMA3-iterative-DPO-final and RLHFlow/LLaMA3-SFT, as in the "Weak-to-Strong Extrapolation Expedites Alignment" paper.

Specifically, we obtain this model by extrapolating (alpha = 0.3) from the weights of the SFT and DPO/RLHF checkpoints, achieving superior alignment with human preference.

Evaluation Results

Evaluation results on the AlpacaEval 2.0 benchmark (you can find the evaluation outputs on the official GitHub repo):

	Win Rate (Ori)	LC Win Rate (Ori)	Win Rate (+ ExPO)	LC Win Rate (+ ExPO)
HuggingFaceH4/zephyr-7b-alpha	6.7%	10.0%	10.6%	13.6%
HuggingFaceH4/zephyr-7b-beta	10.2%	13.2%	11.1%	14.0%
berkeley-nest/Starling-LM-7B-alpha	15.0%	18.3%	18.2%	19.5%
Nexusflow/Starling-LM-7B-beta	26.6%	25.8%	29.6%	26.4%
snorkelai/Snorkel-Mistral-PairRM	24.7%	24.0%	28.8%	26.4%
RLHFlow/LLaMA3-iterative-DPO-final	29.2%	36.0%	32.7%	37.8%
internlm/internlm2-chat-1.8b	3.8%	4.0%	5.2%	4.3%
internlm/internlm2-chat-7b	20.5%	18.3%	28.1%	22.7%
internlm/internlm2-chat-20b	36.1%	24.9%	46.2%	27.2%
allenai/tulu-2-dpo-7b	8.5%	10.2%	11.5%	11.7%
allenai/tulu-2-dpo-13b	11.2%	15.5%	15.6%	17.6%
allenai/tulu-2-dpo-70b	15.4%	21.2%	23.0%	25.7%

Evaluation results on the MT-Bench benchmark (you can find the evaluation outputs on the official GitHub repo):

	Original	+ ExPO
HuggingFaceH4/zephyr-7b-alpha	6.85	6.87
HuggingFaceH4/zephyr-7b-beta	7.02	7.06
berkeley-nest/Starling-LM-7B-alpha	7.82	7.91
Nexusflow/Starling-LM-7B-beta	8.10	8.18
snorkelai/Snorkel-Mistral-PairRM	7.63	7.69
RLHFlow/LLaMA3-iterative-DPO-final	8.08	8.45
internlm/internlm2-chat-1.8b	5.17	5.26
internlm/internlm2-chat-7b	7.72	7.80
internlm/internlm2-chat-20b	8.13	8.26
allenai/tulu-2-dpo-7b	6.35	6.38
allenai/tulu-2-dpo-13b	7.00	7.26
allenai/tulu-2-dpo-70b	7.79	8.03