SeongryongJung/Qwen3-8B-Biology-GRPO-TR

Overview

SeongryongJung/Qwen3-8B-Biology-GRPO-TR is an 8 billion parameter language model derived from the Qwen3-8B base model. It has been fine-tuned using the GRPO (Generalized Reinforcement Learning from Policy Optimization) method, specifically targeting the biological domain. The training utilized a batch size of 32 and was conducted on the SciKnowEval biology dataset.

Key Capabilities & Performance

• Domain Specialization: Optimized for tasks within the biology domain, as evidenced by its training on the SciKnowEval biology dataset.
• Performance Metrics: Achieved a best validation mean@16 score of 50.38% at step 90 on the SciKnowEval biology dataset.
• Context Handling: Supports a maximum prompt length of 2048 tokens and a maximum response length of 8192 tokens, with a maximum model length of 10240 tokens.
• Training Method: Employs the GRPO method, a reinforcement learning approach, for fine-tuning.

Training Details

The model underwent 100 total training steps, with validation performed every 10 steps. Key hyperparameters include a learning rate of 1e-6 and a weight decay of 0.01. The training process involved a train rollout n of 8 and a validation rollout n of 16, utilizing vLLM with 0.8 GPU memory utilization.

Usage

This model is suitable for researchers and developers working on biology-related natural language processing tasks, where specialized domain knowledge is crucial. Its performance on the SciKnowEval biology benchmark suggests its utility in understanding and generating biological text.