While K = 5 (generate 5 samples per iteration), this model uses 3 samples to estimate the soft probabilities P(y_w > y_l) and P(y_l > y_w). These samples include the winner, the loser, and another random sample. This approach has shown to deliver better performance on AlpacaEval 2.0 than the results reported in the paper, but it might also lead to overfitting the PairRM core.
❗Please refer to the original checkpoint at UCLA-AGI/Mistral7B-PairRM-SPPO-Iter3 as reported in our paper. We anticipate that the version in paper demonstrates a more consistent performance improvement across all benchmark tasks.
The following hyperparameters were used during training:
learning_rate: 5e-07
eta: 1000
per_device_train_batch_size: 8
gradient_accumulation_steps: 1
seed: 42
distributed_type: deepspeed_zero3
num_devices: 8
optimizer: RMSProp
lr_scheduler_type: linear
lr_scheduler_warmup_ratio: 0.1
num_train_epochs: 18.0 (stop at epoch=1.0)
Citation
@misc{wu2024self,
title={Self-Play Preference Optimization for Language Model Alignment},
author={Wu, Yue and Sun, Zhiqing and Yuan, Huizhuo and Ji, Kaixuan and Yang, Yiming and Gu, Quanquan},
year={2024},
eprint={2405.00675},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
