Shamika2004/coco_01-ocr-GGUF

coco_01-ocr (GGUF)

coco_01-ocr is a fine-tuned Qwen3-VL-4B model for table OCR. It converts an image of a table into clean HTML (<table>, <tr>, <td>, with rowspan/colspan for merged cells).

It is trained to handle:

• Handwritten text
• Ticks (✓) and crosses (✗)
• Circled or ticked options (☑ selected, ☐ not selected), e.g. ☐ Veg ☑ Non-Veg
• Empty cells kept as <td></td>

Files

File	Description
coco_01-ocr-Q4_K_M.gguf	Language model, 4-bit (smallest, fast)
coco_01-ocr-Q8_0.gguf	Language model, 8-bit (higher quality)
mmproj-coco_01-ocr-f16.gguf	Vision encoder (required, use with any of the above)

Usage (llama.cpp)

llama-mtmd-cli \
  -m coco_01-ocr-Q4_K_M.gguf \
  --mmproj mmproj-coco_01-ocr-f16.gguf \
  --image table.png \
  -p "Convert the table in the image to HTML. Use only <table>, <tr> and <td>, with rowspan/colspan attributes for merged cells. Keep empty cells as <td></td>. Transcribe handwriting exactly. Write ✓ for a tick and ✗ for a cross; keep letters such as Y, N, Yes, No as written. For an option that is circled or ticked (tick next to it, on it, or in a box before it) write ☑ before it and write ☐ before the other option that is not selected."

Works with any llama.cpp-based runner that supports multimodal Qwen3-VL (use a recent build).

Tips

• Crop the image to the table for best results.
• Keep image size around 1 megapixel (the training budget).
• If handwriting accuracy drops with Q4_K_M, use Q8_0.

Training

Fine-tuned with LoRA (TRL SFTTrainer) on synthetic form tables with handwriting, ticks, crosses and circled Yes/No options, plus reviewed real form pages. LoRA weights were merged before GGUF conversion.

Limitations

May make mistakes on very dense, skewed or low-resolution scans. Always verify critical data.