MetaStone-L1-7B
| Property | Value |
|---|---|
| Parameter Count | 7 Billion |
| Base Model | DeepSeek-R1-Distill-Qwen-7B |
| Training Method | GRPO |
| Paper | Graph-Based Synthetic Data Pipeline |
| Model Link | Hugging Face |
What is MetaStone-L1-7B?
MetaStone-L1-7B is a specialized lite reasoning model designed to excel in complex downstream tasks, particularly in mathematics and coding. Built on DeepSeek-R1-Distill-Qwen-7B architecture, it achieves state-of-the-art results among parallel-level models and demonstrates performance comparable to prominent API models like Claude-3.5-Sonnet-1022 and GPT4o-0513.
Implementation Details
The model implementation requires the latest version of transformers (4.48.3) and follows specific optimization guidelines for maximum performance. It utilizes a unique approach with think tags and standardized output formats for different task types.
- Enhanced thoughtful output using think tags
- Standardized input format with user and assistant markers
- Optimized temperature (0.6) and top sampling (0.95)
- Maximum generation length of 32k tokens
Core Capabilities
- Advanced mathematical reasoning with step-by-step solutions
- High-performance code generation and problem-solving
- Structured output formatting for math and coding tasks
- Large context window handling
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its specialized focus on reasoning tasks, particularly in mathematics and coding, achieving SOTA results through its graph-based synthetic data pipeline and optimization techniques.
Q: What are the recommended use cases?
The model excels in mathematical problem-solving, coding tasks, and situations requiring structured reasoning. It's particularly effective when used with standardized prompts for math problems (using \boxed{} for answers) and code generation (using specific formatting guidelines).
