cstr/texo-distill-gguf

Texo-Distill GGUF — Printed Math OCR

GGUF-quantized versions of Texo's distilled PPFormulaNet-S model for on-device LaTeX OCR.

Model Description

Property	Value
Architecture	HGNetv2 (CNN encoder) + MBart (Transformer decoder)
Parameters	20M
Input	384x384 RGB image
Output	LaTeX token sequence
Vocab	1264 tokens (BPE, math-specific)
Encoder	HGNetv2: 4-stage CNN (Conv-BN-ReLU), output 2048-dim
Decoder	MBart: 2 layers, 16 heads, d_model=384, FFN=1536
Training data	UniMER-1M (CC-BY-4.0)
Distilled from	PaddleOCR PP-FormulaNet-S (Apache-2.0)

Performance (UniMER-Test)

Metric	SPE	CPE	SCE	HWE
BLEU	0.9014	0.8909	0.7034	0.8606
Edit distance	0.0780	0.1042	0.1941	0.0995

Available Variants

File	Quant	Size	Quality
texo-distill-f16.gguf	FP16	39 MB	Full precision
texo-distill-q8_0.gguf	Q8_0	22 MB	Identical to FP16
texo-distill-q4_k.gguf	Q4_K	13 MB	Slightly degraded

Recommended: texo-distill-q8_0.gguf (22 MB) — same quality as full precision at 3.6x compression.

Usage with CrispEmbed

# CLI
crispembed -m texo-distill-q8_0.gguf --ocr formula.png

# Auto-detection: the GGUF metadata identifies this as "ppformulanet" architecture
# Works with any CrispEmbed OCR endpoint (CLI, FFI, Flutter)

Image Preprocessing

The model expects UniMERNet-style preprocessing:

1. Convert to grayscale, replicate to 3 channels
2. Resize maintaining aspect ratio to fit 384x384
3. Pad with black (0) to fill 384x384
4. Normalize: mean=0.7931, std=0.1738

License

AGPL-3.0 — inherited from the Texo training code and distilled weights. The original PPFormulaNet-S backbone weights are Apache-2.0 (PaddleOCR), and the training data UniMER-1M is CC-BY-4.0.

Users must accept AGPL-3.0 terms before downloading.

Credits

• Texo by Sicheng Mao — distillation training, architecture design
• PaddleOCR PP-FormulaNet-S — backbone architecture and pretrained weights (Apache-2.0)
• UniMERNet — training dataset and tokenizer (Apache-2.0 / CC-BY-4.0)
• CrispEmbed — GGUF conversion and C++ inference engine

Conversion

Converted using convert-ppformulanet-to-gguf.py from CrispEmbed. BatchNorm layers folded into convolutions. Quantized with crispembed-quantize.