This is a model released for our paper: Regressing the Relative Future: Efficient Policy Optimization for Multi-turn RLHF.

REFUEL-Llama-3-Armo-iter_1

This model is developed with REFUEL based on Meta-Llama-3-8B-Instruct with ArmoRM-Llama3-8B-v0.1 as the reward model and UltraInteract dataset. The training code is available at https://github.com/ZhaolinGao/REFUEL.

Evaluations

Method	Dataset	Winrate at Turn
		h = 1	h = 2	h = 3	h = 4	H = 5	avg
Llama-3.1-70B-it	N/A	70.4	66.4	61.0	53.0	55.4	61.24
REFUEL-Llama-3-Armo-iter_1	REFUEL-Ultrainteract-Llama-3-Armo-iter_1	54.6	53.6	57.8	56.2	59.4	56.32
REFUEL-Llama-3-Armo-iter_2	REFUEL-Ultrainteract-Llama-3-Armo-iter_2	55.2	53.4	58.8	57.2	58.6	56.64

Citation

Please cite our paper if you use this model in your own work:

@misc{gao2024regressingrelativefutureefficient,
      title={Regressing the Relative Future: Efficient Policy Optimization for Multi-turn RLHF}, 
      author={Zhaolin Gao and Wenhao Zhan and Jonathan D. Chang and Gokul Swamy and Kianté Brantley and Jason D. Lee and Wen Sun},
      year={2024},
      eprint={2410.04612},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2410.04612}, 
}