cstr/posformer-mathwriting-GGUF

PosFormer MathWriting — GGUF

GGUF conversion of SJTU-DeepVisionLab/PosFormer for CrispEmbed.

Handwritten mathematical expression OCR — takes a photo of handwritten math and outputs LaTeX. Trained on CROHME 2014.

Architecture

• Encoder: DenseNet (growth_rate=24, 16 layers/block, 1-channel grayscale input)
• Decoder: 3-layer transformer (256-dim, 8 heads, 1024 FFN)
• ARM: coverage-based Attention Reading Module (dc=32) for tracking decoded symbols
• Vocab: 113 tokens, max output length 200
• Input: grayscale image (arbitrary resolution)
• Output: LaTeX string + per-token confidence

Files

File	Size	Description
posformer-mathwriting_v2-f32.gguf	36 MB	F32 full precision

Usage

crispembed --backend posformer -m posformer-mathwriting_v2-f32.gguf \
    --image equation.png

License

BSD-2-Clause (same as the original model).

Credits

• Original model: SJTU-DeepVisionLab
• GGUF conversion & runtime: CrispEmbed