arteeman/mathforge-qwen2.5-math-7b-qlora

MathForge Qwen2.5-Math-7B QLoRA

This repository contains the PEFT/QLoRA adapter for MathForge, an experimental small language model trained to generate competition-math problems from a crux insight and two interacting mathematical techniques.

Architecture

• Base: Qwen/Qwen2.5-Math-7B-Instruct
• Quantization during training: 4-bit NF4
• Adapter: rank-32 LoRA
• Training snapshot: 2,956 integrity-audited examples
• Epochs: 1
• Maximum sequence length: 1,536

This repository contains adapter weights only. The Qwen base weights must be loaded separately.

Evaluation

The current 24-pair exploratory pilot did not show MathForge beating Claude Opus 4.8. See the complete methodology, outputs, and limitations in the GitHub repository:

https://github.com/arteexu/mathforge

Limitations

The model can produce invalid mathematics, incorrect answers, incomplete solutions, and formatting failures. Generated problems require independent mathematical verification before use.