cs-552-2026-thinkinsidethebox/math_model

Overview

This model, developed by cs-552-2026-thinkinsidethebox, is a highly specialized language model designed for advanced mathematical problem-solving. It is a continuation of the previous best OMR-v2 math model, further refined through Supervised Fine-Tuning (SFT).

Key Capabilities

• Mathematical Reasoning: Achieves a pass@8 score of 0.9300 and a pass@1 score of 0.7845 on the local MATH-500 evaluation, indicating strong problem-solving abilities.
• Diagram Interpretation: Enhanced with ASY / diagram-only short-medium OpenMathReasoning repair data, improving its capacity to handle problems with visual components.
• Detailed Explanations: While trained with a "no_think" template, the final exported model utilizes a "thinking" template, suggesting its ability to generate detailed reasoning steps.
• High Extraction Rate: Boasts a boxed_extraction_rate of 0.9305, indicating effective identification of final answers within its responses.

Good For

• Complex Math Problems: Ideal for applications requiring accurate solutions to challenging mathematical questions.
• Educational Tools: Can be integrated into systems that provide step-by-step solutions or explanations for math problems.
• Research in Mathematical AI: Useful for exploring and developing advanced mathematical reasoning capabilities in LLMs.