heli-stand/Qwen2.5-7B-hh-rlhf-Reward

Model Card for Reward Model Fine-tuned on hh-rlhf Dataset from Qwen2.5-7B

This reward model is fine-tuned from Qwen2.5-7B based on the anthropic helpfulness-harmlessness rlhf datasets.

Model Details

Model Description

This is a fine-tuned Qwen2.5-7B model, trained on the anthropic helpfulness-harmlessness rlhf datasets.

• —Model type: Transformer-based language model
• —Language(s) (NLP): English
• —License: apache-2.0
• —Finetuned from model: Qwen2.5-7B

Model Sources

• —Repository: https://huggingface.co/Qwen/Qwen2.5-7B

Training Procedure

Training Data

• —Repository: https://huggingface.co/datasets/heli-stand/hh-rlhf

Training Hyperparameters

• —Optimizer: Adamw
• —Learning Rate: 5e-6, search range: 3e-6, 5e-6, 1e-5
• —Batch Size: 128
• —Epochs: 1
• —Max_Length: 1024

Evaluation Results

• —Best Evaluation Loss: 0.5539
• —Best Evaluation Accuracy: 0.7094