Light-R1-7B-DS
| Property | Value |
|---|---|
| Author | qihoo360 |
| Base Model | DeepSeek-R1-Distill-Qwen-7B |
| Model Size | 7B parameters |
| Release Date | March 12, 2025 |
| HuggingFace | Link |
What is Light-R1-7B-DS?
Light-R1-7B-DS represents a significant breakthrough in mathematical reasoning AI models, achieving state-of-the-art performance with remarkably efficient training. Built upon DeepSeek-R1-Distill-Qwen-7B, this model has been fine-tuned with just 3,000 carefully curated training examples, demonstrating exceptional performance on challenging mathematical benchmarks.
Implementation Details
The model builds upon the DeepSeek-R1-Distill-Qwen-7B architecture and implements rigorous data decontamination practices, including exact matching (excluding digits) and N-gram (N=32) matching to ensure benchmark integrity. It maintains the same usage pattern as its base model while achieving superior performance.
- Achieves 59.1% accuracy on AIME24 and 44.3% on AIME25
- 49.4% performance on GPQA without specific training
- Trained with only 3K SFT data
- Implements thorough data decontamination protocols
Core Capabilities
- Advanced mathematical reasoning and problem-solving
- Strong performance on American Invitational Mathematics Examination (AIME) problems
- Generalized problem-solving capabilities demonstrated through GPQA performance
- Efficient learning from limited training data
Frequently Asked Questions
Q: What makes this model unique?
Light-R1-7B-DS stands out for achieving state-of-the-art performance on mathematical reasoning tasks with minimal training data (3K examples), demonstrating exceptional efficiency in learning and generalization.
Q: What are the recommended use cases?
The model is particularly suited for mathematical problem-solving, especially in competitive mathematics contexts and general mathematical reasoning tasks. It shows strong capabilities in both specialized mathematics (AIME) and general problem-solving (GPQA).
