Self-Play Preference Optimization for Language Model Alignment (https://arxiv.org/abs/2405.00675)

Mistral7B-PairRM-SPPO-Iter1

This model was developed using Self-Play Preference Optimization at iteration 1, based on the mistralai/Mistral-7B-Instruct-v0.2 architecture as starting point. We utilized the prompt sets from the openbmb/UltraFeedback dataset, splited to 3 parts for 3 iterations by snorkelai/Snorkel-Mistral-PairRM-DPO-Dataset. All responses used are synthetic.

This is the model reported in the paper , with K=5 (generate 5 responses per iteration). We attached the Arena-Hard eval results in this model page.

Links to Other Models

• Mistral7B-PairRM-SPPO-Iter1
• Mistral7B-PairRM-SPPO-Iter2
• Mistral7B-PairRM-SPPO-Iter3
• Mistral7B-PairRM-SPPO

Model Description

• Model type: A 7B parameter GPT-like model fine-tuned on synthetic datasets.
• Language(s) (NLP): Primarily English
• License: Apache-2.0
• Finetuned from model: mistralai/Mistral-7B-Instruct-v0.2

AlpacaEval Leaderboard Evaluation Results

Model	LC. Win Rate	Win Rate	Avg. Length
Mistral7B-PairRM-SPPO Iter 1	24.79	23.51	1855
Mistral7B-PairRM-SPPO Iter 2	26.89	27.62	2019
Mistral7B-PairRM-SPPO Iter 3	28.53	31.02	2163
Mistral7B-PairRM-SPPO Iter 1 (best-of-16)	28.71	27.77	1901
Mistral7B-PairRM-SPPO Iter 2 (best-of-16)	31.23	32.12	2035
Mistral7B-PairRM-SPPO Iter 3 (best-of-16)	32.13	34.94	2174

Arena-Hard Evaluation Results

Model	Score	95% CI	average # Tokens
Mistral7B-PairRM-SPPO-Iter3	23.3	(-1.8, 1.8)	578

Open LLM Leaderboard Evaluation Results

Results are reported by using lm-evaluation-harness v0.4.1

	arc_challenge	truthfulqa_mc2	winogrande	gsm8k	hellaswag	mmlu	average
Mistral7B-PairRM-SPPO Iter 1	65.02	69.4	77.82	43.82	85.11	58.84	66.67
Mistral7B-PairRM-SPPO Iter 2	65.53	69.55	77.03	44.35	85.29	58.72	66.75
Mistral7B-PairRM-SPPO Iter 3	65.36	69.97	76.8	42.68	85.16	58.45	66.4

MT-Bench Evaluation Results

	1st Turn	2nd Turn	Average
Mistral7B-PairRM-SPPO Iter 1	7.63	6.79	7.21
Mistral7B-PairRM-SPPO Iter 2	7.90	7.08	7.49
Mistral7B-PairRM-SPPO Iter 3	7.84	7.34	7.59

Training hyperparameters

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}
}