Hugging Face Overview
Overview
TIGER-Lab/Qwen2.5-Math-7B-CFT is a 7.6 billion parameter model specifically designed for mathematical reasoning. It introduces a novel Critique Fine-Tuning (CFT) methodology, which trains the model to critique and analyze responses, fostering a deeper understanding of mathematical concepts rather than simply imitating correct answers. This approach is inspired by human learning processes that emphasize critical thinking.
Key Capabilities
- Novel Training Paradigm: Employs Critique Fine-Tuning (CFT) for enhanced reasoning, moving beyond traditional supervised fine-tuning (SFT).
- Exceptional Data Efficiency: Achieves performance comparable to models trained on 40 times more data, utilizing only 50K samples from the WebInstruct-CFT-50K dataset.
- Superior Mathematical Performance: Demonstrates consistent 4-10% improvement over traditional SFT methods across six math benchmarks.
- High Accuracy: Reaches 79.4% accuracy on the MATH benchmark and 41.6% on OlympiadBench, matching or exceeding models with significantly larger training datasets.
Good For
- Mathematical Reasoning Tasks: Ideal for complex math problems, including those found in competitive programming and academic challenges.
- Research in LLM Training: Provides a strong example of an alternative, highly effective fine-tuning strategy for specialized domains.
- Applications Requiring Critical Analysis: Suitable for scenarios where a model needs to not just provide an answer, but also understand and evaluate the reasoning behind it.