University-HMER-UniMuMER-LoRA

University-HMER-UniMuMER-LoRA

This repository contains the Uni-MuMER LoRA adapter used in the University HMER project.

Model description

The adapter fine-tunes phxember/Uni-MuMER-Qwen3.5-2B on a real classroom handwritten calculus dataset.

Training used LoRA with Unsloth.

Intended role

This is the main robust model used for the project demo.

It is designed for:

• real classroom handwritten formula images;
• phone-camera inputs;
• imperfect lighting and background conditions;
• broader live-demo robustness compared with the compact TAMER specialist model.

Dataset

Dataset: University-HMER-RealClassroom

Real classroom split:

train:       1103 samples
validation:  259 samples
blind test:  274 samples
total:      1636 images

Adapter files

Uploaded adapter directory:

outputs/unimumer_lora_unsloth_real/best_adapter/

The repository contains the LoRA adapter files and tokenizer/processor assets needed for inference with the base Uni-MuMER model.

Results

Validation

Blind test

Inference note

Load the base model first:

phxember/Uni-MuMER-Qwen3.5-2B

Then load this repository as a LoRA adapter.

Limitations

The adapter is trained on a relatively small real classroom dataset. It is more robust than the compact TAMER specialist model for live inputs, but it can still fail on unusual symbols, ambiguous handwriting, severe blur, or expressions outside the training distribution.

Acknowledgements

Thanks to:

• Uni-MuMER: https://github.com/BFlameSwift/Uni-MuMER
• Unsloth: https://github.com/unslothai/unsloth