---
base_model: Qwen/Qwen3-VL-4B-Instruct
library_name: gguf
pipeline_tag: image-text-to-text
tags:
  - gguf
  - qwen3-vl
  - ocr
  - table-recognition
  - llama.cpp
---

# coco1-ocr (GGUF)

Fine-tuned from [Qwen/Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct) (LoRA, merged) for document understanding: table-to-HTML, checkbox/option transcription, handwritten line transcription, digit recognition, and layout detection.

## Files

| File | Purpose |
|---|---|
| `coco1-ocr-Q8_0.gguf` | Language model (Q8_0) |
| `mmproj-coco1-ocr-f16.gguf` | Vision projector (F16), required for image input |

## Usage (llama.cpp)

```bash
llama-mtmd-cli -m coco1-ocr-Q8_0.gguf --mmproj mmproj-coco1-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>."
```

## Prompts used in training

- **table:** 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>.
- **check_text:** Transcribe the options shown. Put ☑ before the option that is marked or selected and ☐ before every other option.
- **hw_line:** Transcribe the handwritten text.
- **digit:** What digit is shown? Answer with the digit only.
- **layout:** List every layout element with its category and bounding box [x0,y0,x1,y1] in 0-1000 relative coordinates, from top to bottom.

## Notes

Converted to GGUF from the merged fine-tuned weights. Use the same prompts as above for best results.