cstr/trocr-math-handwritten-gguf

TrOCR Math Handwritten — GGUF

Handwritten math equation recognition via CrispEmbed's ggml inference engine.

Source model: fhswf/TrOCR_Math_handwritten (AFL-3.0)

Architecture: ViT-Large encoder (24L, 1024d, 16H) + TrOCR-Large decoder (12L, 1024d, 16H, vocab=50265)

Models

File	Size	Description
trocr-math-handwritten-fp16.gguf	1.2 GB	FP16 weights
trocr-math-handwritten-q4_k.gguf	355 MB	Q4_K quantized

Verified Output

Input	Recognized
x+y=5 (hand-drawn)	X + Y =
x²+1 (printed)	x^{2}
3x+7=0 (printed)	3x+7
+ (cross pattern)	+

Usage with CrispEmbed

math_ocr_context* ctx = math_ocr_init("trocr-math-handwritten-q4_k.gguf", 4);
const char* latex = math_ocr_recognize(ctx, gray_pixels, width, height, &len);
math_ocr_free(ctx);

Conversion

Converted from HuggingFace safetensors using CrispEmbed's torch-free converter:

python models/convert-trocr-safetensors-to-gguf.py \
  --model-dir /path/to/fhswf/TrOCR_Math_handwritten \
  --output trocr-math-handwritten-fp16.gguf --fp16

Key converter features:

• No PyTorch dependency (uses safetensors + gguf + numpy only)
• Generates sinusoidal position embeddings (not stored in checkpoint)
• Squeezes 3D batch dimensions, reshapes 4D conv weights
• Preserves AFL-3.0 license in GGUF metadata

License

Academic Free License v3.0 (inherited from source model)

Provenance and EU AI Act Art. 53 note

• Upstream model: fhswf/TrOCR_Math_handwritten — published by fhswf.
• Upstream licence: afl-3.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.