GLM-OCR EOT Fix (Q8_0) for Ollama

A community repackaging of Ollama’s glm-ocr:q8_0 with a GGUF end-of-turn metadata adjustment intended to prevent runaway and repetitive generation during OCR.

This is an unofficial community build. It is not affiliated with or endorsed by Z.ai or Ollama.

What changed?

• Based on the glm-ocr:q8_0 model distributed through the Ollama Library.
• Repacked the GGUF with gguf-new-metadata --special-token eot '<|user|>' to set the end-of-turn token metadata.
• The model was not retrained; this is a metadata-level change, not a fine-tune.
• Uses temperature 0 in the Ollama Modelfile.

The change is intended to mitigate runaway generation and repeated output reported by some users when GLM-OCR is used through Ollama. It is a workaround, not a guarantee that every generation issue is fixed on every setup.

Install

ollama pull mhruiz/glm-ocr-eotfix:latest

Usage

Run interactively:

ollama run mhruiz/glm-ocr-eotfix:latest

Example OCR request for a local image:

Text Recognition: ./image.png

Validation

Tested locally by the publisher with documents that had triggered repeated OCR output in the original Ollama build. Results may vary with Ollama versions, prompt formats and input documents.

Upstream and background

• Upstream GLM-OCR model card: https://huggingface.co/zai-org/GLM-OCR
• Original Ollama model: https://ollama.com/library/glm-ocr
• Ollama issue discussing repeated generation: https://github.com/ollama/ollama/issues/18609
• Related Ollama pull request: https://github.com/ollama/ollama/pull/17195

License and attribution

The upstream GLM-OCR model card declares the model weights under the MIT License. See the accompanying LICENSE file.

This repository/model is an unofficial repackaging with a metadata change. It does not imply endorsement by Z.ai or Ollama. Consult the upstream model card for authoritative licensing and attribution information.