XingYing-stack/TIPS-Qwen3-4B-Thinking-2507-Math
XingYing-stack/TIPS-Qwen3-4B-Thinking-2507-Math is a 4 billion parameter generative reward model, initialized from Qwen/Qwen3-4B-Thinking-2507 and trained with outcome-only GRPO. This model specializes in reasoning over mathematical solutions to produce step-level and outcome labels, rather than general-purpose chat. It is designed for reward modeling and process verification in mathematical contexts, leveraging a 32768 token context length.
Loading preview...
Model Overview
The XingYing-stack/TIPS-Qwen3-4B-Thinking-2507-Math is a 4 billion parameter generative reward model. It is built upon the Qwen/Qwen3-4B-Thinking-2507 base model and further trained using outcome-only Generative Reward Policy Optimization (GRPO). This model's core function is to analyze mathematical solutions, providing detailed step-level and final outcome labels.
Key Capabilities
- Mathematical Reasoning: Excels at reasoning through mathematical solutions.
- Generative Reward Modeling: Produces step-level and outcome labels for mathematical problems.
- Process Verification: Designed for verifying the steps and results within mathematical problem-solving.
- High Context Length: Supports a context length of 32768 tokens, allowing for analysis of complex problems.
Good For
- Reward Modeling: Ideal for applications requiring a reward signal based on the correctness of mathematical steps and outcomes.
- Educational Tools: Can be used in systems that provide feedback on mathematical problem-solving processes.
- Automated Grading: Potentially useful for automated evaluation of mathematical solutions by verifying intermediate steps.
This model is specifically intended for reward modeling and process verification in mathematical domains, not for general-purpose conversational AI. Training data and code are available via Hugging Face Datasets and GitHub.