Ba2han/math-test-maxx
Ba2han/math-test-maxx is a 3.1 billion parameter language model based on the Qwen2.5-3B-Instruct architecture, developed by Ba2han. This model is specifically optimized for mathematical reasoning tasks, achieving improved performance on benchmarks like GSM8K and MATH-500 through a specialized ReasonMaxxer offline-search LoRA. It is designed to enhance mathematical problem-solving capabilities over its base model.
Loading preview...
Model Overview
Ba2han/math-test-maxx is a 3.1 billion parameter model built upon the unsloth/Qwen2.5-3B-Instruct architecture. It incorporates a ReasonMaxxer offline-search LoRA (v2 recipe) to enhance its mathematical reasoning abilities. The model was trained using a specific recipe involving a LoRA configuration (r=16, α=32, QKVO), a learning rate of 2e-5, and a max gradient norm of 0.1. The training data included 300 MATH-500 items and 12 offline search samples, utilizing an entropy-weighted signed loss function over 774 micro-steps.
Key Capabilities
- Enhanced Mathematical Reasoning: Demonstrates improved performance on mathematical benchmarks compared to its base model.
- Specific LoRA Integration: Utilizes a ReasonMaxxer LoRA for targeted optimization in mathematical problem-solving.
Performance Highlights
Evaluated with vLLM, using a boxed chat prompt and MathVerifier (0-shot greedy):
- GSM8K (n=1319): Achieved 85.0%, a +0.4 percentage point improvement over the base model (84.6%).
- MATH-500 (n=500): Achieved 62.8%, a +1.2 percentage point improvement over the base model (61.6%).
Good For
- Applications requiring strong mathematical reasoning and problem-solving.
- Tasks where incremental improvements on math benchmarks are critical.
- Developers looking for a fine-tuned Qwen2.5-3B-Instruct variant with a focus on quantitative tasks.