Qwen3 4B LaTeX Correction v3 BF16
This repository contains the full BF16 merged checkpoint for the Qwen3 4B LaTeX correction SFT release.
The model restores noisy, colloquial, or pseudocode-like math descriptions into canonical LaTeX formulas. It is trained to return formula-only output in a strict display dialect: fractions use the standard LaTeX frac command with two braced arguments, matrices use the bmatrix environment, and superscripts/subscripts are braced, for example x_{i} and x^{2}.
Recommended Use
Use this BF16 repository for Transformers-compatible inference, vLLM-compatible conversion, continued fine-tuning, or producing new quantizations.
A local-inference GGUF build is published separately:
Evaluation
Clean evaluation split: 292 examples.
| model | exact accuracy | semantic accuracy | parse coverage |
|---|
| Base Qwen3-4B-Instruct-2507 | 2.40% | 13.36% | 36.99% |
| Qwen3-4B LaTeX Correction LoRA v3 | 83.22% | 98.97% | 100.00% |
semantic_accuracy is computed by the repository's local LaTeX AST matcher and is the primary metric.
Source
