SuperCorrect-7B
| Property | Value |
|---|---|
| Parameter Count | 7.62B |
| Base Model | Qwen2.5-Math-7B-Instruct |
| License | Apache-2.0 |
| Paper | arxiv:2410.09008 |
| Tensor Type | BF16 |
What is SuperCorrect-7B?
SuperCorrect-7B is a state-of-the-art language model specifically designed for mathematical reasoning tasks. Built on the Qwen2.5-Math-7B architecture, it implements a novel two-stage fine-tuning method that enhances both reasoning accuracy and self-correction capabilities. The model has demonstrated remarkable performance, outperforming DeepSeekMath-7B by 7.8%/5.3% and Qwen2.5-Math-7B by 15.1%/6.3% on MATH/GSM8K benchmarks.
Implementation Details
The model utilizes a pre-defined hierarchical thought template called Buffer of Thought (BoT) for more deliberate reasoning compared to conventional Chain-of-Thought approaches. It processes mathematical problems through XML-formatted steps, with special attention to challenging aspects that require detailed explanations.
- Implements error-driven insights for improved accuracy
- Uses XML-based structured reasoning format
- Requires transformers >= 4.37.0
- Supports both CPU and GPU inference
Core Capabilities
- Advanced mathematical problem-solving
- Step-by-step reasoning with self-correction
- Hierarchical thought processing
- Detailed explanation generation for complex steps
- Performance optimization for mathematical benchmarks
Frequently Asked Questions
Q: What makes this model unique?
SuperCorrect-7B stands out for its innovative two-stage fine-tuning approach and the incorporation of hierarchical thought templates. Unlike other models, it focuses on pure mathematical reasoning abilities without relying on programming methods like PoT or ToRA.
Q: What are the recommended use cases?
The model excels in mathematical problem-solving scenarios, particularly in educational contexts where step-by-step reasoning and detailed explanations are valuable. It's ideal for tackling complex mathematical problems that require structured thinking and self-correction capabilities.
