AdarshSingh7647/Eklav-4B-Math
AdarshSingh7647/Eklav-4B-Math is a 4 billion parameter language model based on Qwen/Qwen3-4B, specifically optimized for mathematical reasoning tasks. It utilizes a unique 'Eklav' training method, where the model learns to continue reasoning from partial teacher traces rather than imitating full reasoning paths. This approach enhances its ability to solve complex math problems, making it suitable for applications requiring robust mathematical problem-solving capabilities.
Loading preview...
Eklav-4B-Math: Enhanced Mathematical Reasoning
AdarshSingh7647/Eklav-4B-Math is a 4 billion parameter model built upon the Qwen/Qwen3-4B base, distinguished by its innovative 'Eklav' training methodology. Unlike standard Chain-of-Thought (CoT) distillation, Eklav trains the model to complete a teacher's reasoning process from a partial trace, rather than reproducing the entire sequence. This method focuses on developing the model's ability to independently continue reasoning and derive answers.
Key Capabilities
- Advanced Math Reasoning: The Eklav training method significantly improves the model's performance on various mathematical benchmarks.
- Efficient Learning: It learns to infer and complete reasoning steps from hints, fostering a deeper understanding of problem-solving.
- Performance Gains: Achieves an average +3.5% pass@1 across 6 math benchmarks compared to standard full trace CoT SFT, using the same base model and training data.
Benchmarks
The model demonstrates superior performance in math-specific evaluations:
- AIME 2025: 36.0% (vs. 31.9% for CotGen)
- GSM8K: 94.3% (vs. 93.8% for CotGen)
- MATH-500: 91.0% (vs. 88.7% for CotGen)
- Omni-MATH: 38.2% (vs. 34.7% for CotGen)
Good For
- Applications requiring strong mathematical problem-solving.
- Educational tools for generating step-by-step math solutions.
- Research into advanced reasoning and learning from partial information.