PosFormer Handwritten Math OCR — CROHME-trained GGUF
PosFormer (Position-aware Transformer) for handwritten mathematical
expression recognition,
retrained from scratch on CROHME 2014 and
converted to GGUF format for
CrispEmbed.
License
CC BY-NC-SA 3.0 — inherited from the CROHME 2014 training data.
- Personal, educational, and research use: allowed
- Apps (including commercial): allowed — the app is a separate work;
the NC clause applies to the weights, not to software that loads them.
Users download the weights separately and accept the NC terms.
- Redistributing or selling the weights themselves: not allowed
- Attribution required: cite CROHME and PosFormer (see below)
- ShareAlike: derivative models must use the same or compatible license
The C++ inference engine (
CrispEmbed)
and GGUF converter are original clean-room implementations (MIT license).
Model details
| Property | Value |
|---|
| Architecture | DenseNet encoder + 3-layer Transformer decoder + ARM |
| Parameters | 6.5M |
| Training data | CROHME 2014 (8,835 images) + MathWriting (2,000 images) |
| Training data license | CC BY-NC-SA 3.0 (CROHME) + CC BY-NC-SA 4.0 (MathWriting) |
| Training | 300 epochs, SGD+momentum, cosine annealing warm restarts, label smoothing 0.1 |
| Vocabulary | 113 LaTeX tokens (canonical PosFormer dictionary) |
| Input | Grayscale handwritten math image |
| Output | LaTeX token sequence |
Training details
Retrained from scratch (no transfer learning from published weights) using
the PosFormer architecture on CROHME 2014 train set. Key differences from
the published training:
- Cosine annealing with warm restarts (T_0=30, T_mult=2) instead of ReduceLROnPlateau
- Label smoothing (ε=0.1) on cross-entropy loss
- Greedy validation (beam_size=1) for faster training epochs
Training monitored via Weights & Biases.
Files
| File | Quant | Size | Notes |
|---|
posformer-crohme-f32.gguf | F32 | 24.9 MB | Full precision |
posformer-crohme-q8_0.gguf | Q8_0 | 12 MB | Recommended for mobile |
posformer-crohme-q4_k.gguf | Q4_K | 10 MB | Smallest, lossless on test |
Accuracy (CROHME 2014 test set, 986 images)
Greedy left-to-right decoding (beam_size=1, no bi-directional search):
| Model | Epoch | Raw match | Notes |
|---|
| This model (F32) | 182 | 60.5% | Retrained, CC BY-NC-SA 3.0/4.0 |
| SJTU published weights | 206 | 56.0% | Academic-only license |
| BTTR baseline | — | 49.2% | MIT license |
| HMER baseline | — | 36.1% | MIT license |
Trained on CROHME 2014 (8835 images) + 2000 MathWriting samples (filtered
to 110-token vocab). 60.5% beam=1 exact match on full CROHME 2014 test
(986 images). val_ExpRate peaked at 62.0% during training. Surpasses
published SJTU greedy (56.0%) by 4.5 points. LR=0.00125 after
ReduceLROnPlateau drop.
Note: the published PosFormer ExpRate of 62.7% uses bi-directional beam
search (beam_size=10). Our greedy results are directly comparable to our
C++ inference engine which uses greedy L2R decoding.
Usage with CrispEmbed
1# Build
2cd CrispEmbed-build
3cmake /path/to/CrispEmbed
4make -j$(nproc) test-posformer
5
6# Run
7export LD_LIBRARY_PATH=$PWD/ggml/src
8./test-posformer posformer-crohme-q8_0.gguf image.bmp
Parity
The C++ inference engine matches PyTorch to >99.999% (cosine similarity
= 1.000000 at every decoder step). See tests/parity/posformer_*.py
in the CrispEmbed repo for verification scripts.
Citation
1@inproceedings{chen2024posformer,
2 title={PosFormer: Recognizing Complex Handwritten Mathematical Expression
3 with Position Forest Transformer},
4 author={Chen, Tongkun and others},
5 booktitle={AAAI},
6 year={2024}
7}
References
Provenance — authored here, and outside the GPAI definition
Added 2026-08-02 during an account-wide provenance review.
These weights were trained by this repository's maintainer, not converted
from someone else's model. Most cstr/* repositories are GGUF/ONNX conversions
where the upstream research team remains the provider; this one is not, and the
distinction matters because the two attract different obligations. The training
data, hyperparameters and licence inheritance are documented above.
EU AI Act Art. 53 does not apply. Art. 53 binds providers of general-purpose
AI models, which Art. 3(63) defines as models displaying "significant generality"
and capable of "competently performing a wide range of distinct tasks". PosFormer recognising handwritten mathematical expressions
is a narrow, single-task model and does not meet that definition, so the
Art. 53(1)(c) copyright-policy and 53(1)(d) training-content duties are not
engaged. This is recorded explicitly because "trained here" and "subject to
Art. 53" are easy to conflate, and only the first is true.
The training-data documentation above is published because it is useful and
because the licence inheritance depends on it — not because Art. 53 compels it.
