amildravid4292/Qwen3-4B-Base-GRPO-MATH-ck300
amildravid4292/Qwen3-4B-Base-GRPO-MATH-ck300 is a 4 billion parameter Qwen3-Base model fine-tuned using GRPO (RL with a verifiable correctness reward) specifically on the MATH dataset. This model is optimized for mathematical reasoning tasks, demonstrating strong performance on the MATH-500 benchmark with a pass@1 score of 0.767. It is designed for raw text evaluation under a boxed prompt format, making it suitable for complex mathematical problem-solving.
Loading preview...
Model Overview
amildravid4292/Qwen3-4B-Base-GRPO-MATH-ck300 is a specialized 4 billion parameter Qwen3-Base model. It has undergone 300 steps of GRPO (RL with a verifiable correctness reward) fine-tuning on the MATH training dataset, with a rank-64 LoRA adapter merged into the base weights. This model is intended for raw text evaluation using a specific boxed prompt format, without the need for a chat template.
Key Capabilities
- Mathematical Reasoning: Achieves a pass@1 score of 0.767 on the held-out MATH-500 dataset, indicating strong performance in solving complex math problems.
- Reinforcement Learning Fine-tuning: Utilizes GRPO with a binary correctness reward for the final boxed answer, enhancing its ability to produce verifiable solutions.
- Prompt Format: Designed to be evaluated with a specific prompt structure:
Problem: {q}\nPlease reason step by step, and put your final answer within \boxed{}.\nSolution: - AIME Performance: Shows a pass@1 of 0.114 and pass@32 of 0.333 on AIME 2025 problems.
When to Use This Model
This model is particularly well-suited for:
- Mathematical Problem Solving: Ideal for applications requiring step-by-step reasoning and accurate final answers in mathematical contexts.
- Research in RL for Math: Useful for researchers exploring reinforcement learning techniques for improving mathematical capabilities in LLMs.
- Benchmarking: Can serve as a strong baseline or comparison point for other models focused on the MATH dataset.