| Property | Value |
|---|---|
| Base Model | Mistral-7B |
| Paper | WizardMath Paper |
| Task Type | Mathematical Reasoning |
| Performance | 83.2% GSM8k, 33.0% MATH |
What is WizardMath-7B-V1.1?
WizardMath-7B-V1.1 is a cutting-edge mathematical reasoning language model developed by the WizardLM team. Built on the Mistral-7B architecture, it represents a significant breakthrough in mathematical problem-solving capabilities for smaller language models, outperforming many larger models including ChatGPT 3.5 and Gemini Pro.
Implementation Details
The model utilizes the Reinforced Evol-Instruct (RLEIF) methodology and is specifically optimized for mathematical reasoning tasks. It requires specific system prompts for optimal performance and offers both default and Chain-of-Thought (CoT) versions for different complexity levels of mathematical problems.
- Built on Mistral-7B architecture
- Implements Reinforced Evol-Instruct methodology
- Rigorously tested for data contamination
- Supports both standard and CoT prompting
Core Capabilities
- Achieves 83.2% pass@1 on GSM8k benchmark
- Scores 33.0% pass@1 on MATH dataset
- Outperforms larger models like Mixtral MOE (74.4% GSM8k)
- Excels in step-by-step mathematical reasoning
Frequently Asked Questions
Q: What makes this model unique?
WizardMath-7B-V1.1 achieves unprecedented performance for its size, surpassing much larger models while maintaining efficiency. It's particularly notable for achieving state-of-the-art results in the 7B parameter category for mathematical reasoning tasks.
Q: What are the recommended use cases?
The model is optimized for mathematical problem-solving, particularly suited for educational applications, tutoring systems, and automated math problem verification. For simple math questions, the default prompt is recommended, while complex problems benefit from the Chain-of-Thought prompt.