RefAlign: RL with Similarity-based Rewards

GitHub repository: https://github.com/mzhaoshuai/RefAlign

Paper: Learning from Reference Answers: Versatile Language Model Alignment without Binary Human Preference Data.

This is the model aligned with SimPO described in the paper Learning from Reference Answers: Versatile Language Model Alignment without Binary Human Preference Data.

This model is a fine-tuned version of meta-llama/Meta-Llama-3-8B-Instruct on the mzhaoshuai/llama3-ultrafeedback-bertscore-bart-large-mnli dataset. It achieves the following results on the evaluation set:

• Loss: 1.3795
• Rewards/chosen: -9.7179
• Rewards/rejected: -15.1619
• Rewards/accuracies: 0.8770
• Rewards/margins: 5.4440
• Logps/rejected: -1.5162
• Logps/chosen: -0.9718
• Logits/rejected: -1.3590
• Logits/chosen: -1.2077

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-06
• train_batch_size: 2
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 16
• total_train_batch_size: 128
• total_eval_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Framework versions

• Transformers 4.44.2
• Pytorch 2.2.2+cu121
• Datasets 4.0.0
• Tokenizers 0.19.1