cstr/hmer-handwritten-math-gguf

HMER Handwritten Math OCR — GGUF

HMER (Handwritten Mathematical Expression Recognition) with DenseNet-121 encoder and GRU attention decoder, converted to GGUF format for CrispEmbed.

License

CC BY-NC-SA 3.0 — the model code is MIT (whywhs/Pytorch-HMER), but the weights were trained on CROHME 2016 (CC BY-NC-SA 3.0). The training data license propagates to the trained weights.

• Personal, educational, and research use: allowed
• Apps (including commercial): allowed — 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

The C++ inference engine (CrispEmbed, MIT license) and GGUF converter are original clean-room implementations.

Model details

Property	Value
Architecture	DenseNet-121 encoder + GRU attention decoder (with coverage)
Parameters	~6.8M
Training data	CROHME 2016 (CC BY-NC-SA 3.0)
Vocabulary	112 LaTeX tokens
CROHME 2014 eval	36.1% raw / 36.3% parsed (greedy, 986 images)

Files

File	Quant	Size
hmer-hw-f32.gguf	F32	~14 MB
hmer-hw-f16.gguf	F16	~7 MB
hmer-hw-q8_0.gguf	Q8_0	7 MB
hmer-hw-q4_k.gguf	Q4_K	4 MB

References

• Pytorch-HMER code (MIT)
• CrispEmbed inference engine (MIT)
• CROHME competition