nakamura196/yigdzin1-gguf

Yigdzin-1 — GGUF (llama.cpp)

GGUF conversion of BDRC/tibetan-ocr ("Yigdzin-1"), a Tibetan OCR vision-language model fine-tuned from PaddlePaddle/PaddleOCR-VL-1.6, for use with llama.cpp / llama-server (mtmd multimodal support). Used by the local-ocr desktop app.

Files

• yigdzin1.gguf — the language model (Q8_0-equivalent precision inherited from the source checkpoint's conversion; same conversion path as PaddlePaddle/PaddleOCR-VL-1.6-GGUF)
• yigdzin1-mmproj.gguf — the vision projector (--mmproj)

Conversion notes

Converted with a stock (unmodified) llama.cpp build (b10776) using its conversion/ernie.py PaddleOCR-VL converter, with one compatibility fix: this checkpoint ships processor_config.json with size.shortest_edge/size.longest_edge instead of the flat min_pixels/max_pixels keys the base PaddleOCR-VL-1.6 uses, which the stock converter doesn't handle. No change was made to inference-time behavior — the produced GGUF carries no non-standard metadata and loads/runs identically to any other PaddleOCR-VL-arch GGUF on stock llama-server.

(An earlier, separate experiment explored whether Yigdzin-1's image-token positional encoding needs a non-standard "sequential" mode at inference time — relevant for whole-page images. It was found unnecessary for the line-level crops this conversion targets; see the source app's repository for details.)

Usage

llama-server -m yigdzin1.gguf --mmproj yigdzin1-mmproj.gguf

Then send a chat-completion request with an image (line crop) and the prompt "Extract all Tibetan text.".

License

Apache-2.0, inherited from BDRC/tibetan-ocr.