University-HMER-TAMER-A3-RealFT
This repository contains the TAMER-A3 RealFT checkpoint used in the University HMER project.
Model description
TAMER-A3 RealFT is a fast specialist HMER model adapted for handwritten university-level calculus expressions collected from real classroom-style images.
The model follows a TAMER-based architecture with dual adapter components:
- encoder adapter enabled;
- decoder adapter enabled;
- fine-tuned on the real classroom dataset after the University12K adaptation stage.
Intended role
This model is intended as a fast specialist baseline for the collected real-classroom distribution.
It is useful for:
- domain adaptation analysis;
- speed/accuracy trade-off comparison;
- research comparison against Uni-MuMER LoRA;
- reproducing TAMER-A3 RealFT results.
For unconstrained live demo input, the Uni-MuMER LoRA model is recommended as the primary robust model.
Dataset
Dataset: University-HMER-RealClassroom
Real classroom split:
train: 1103 samples
validation: 259 samples
blind test: 274 samples
total: 1636 images
Checkpoint
Uploaded checkpoint:
epoch=56-val_university_ExpRate=0.5637.ckpt
Expected local path in the original research project:
outputs/real_ft_a3_dual_seed7/checkpoints/epoch=56-val_university_ExpRate=0.5637.ckpt
Results
Validation
| Model | ExpRate | TER | Latency |
|---|---|---|---|
| TAMER-A3 RealFT | 56.37% | 5.45% | 0.299 s/img |
Blind test
| Model | ExpRate | TER | Latency |
|---|---|---|---|
| TAMER-A3 RealFT | 71.17% | 2.92% | 0.306 s/img |
Limitations
The model is specialized for the collected real classroom distribution. It may be less robust for free-form live inputs with different writing styles, lighting, cropping, or spatial layouts.
A known limitation of the real dataset is reduced diversity in operator-bound placement. Many integral and summation bounds are written diagonally or to the right of the operator, while vertically stacked display-style bounds are underrepresented.
Citation / acknowledgement
This work extends the original TAMER implementation: