OpenGlyph-OCR-0 / NOTICE
astromandarin's picture
OpenGlyph-OCR-0: the model in two forms (transformers and MLX), model card, licence
853865d verified
Raw History Blame Contribute Delete
2.16 kB
OpenGlyph-OCR-0
This repository contains, in base/, the model Xiaomi-OCR-0
(https://huggingface.co/xiaomi-research/Xiaomi-OCR-0), released by its authors under the
Apache License, Version 2.0, and itself built on Qwen3.5-0.8B-Base. The files in base/
are redistributed unchanged; their SHA-256 hashes are listed in release.json.
model.safetensors is a derivative of that model: its weights with the changes named
below applied. config.json, generation_config.json, tokenizer.json,
tokenizer_config.json, processor_config.json and chat_template.jinja are the files of
Xiaomi-OCR-0 with what those changes need entered (the added tokens, the larger
vocabulary, the output layer of this model); chat_template.jinja is unchanged.
The other files -- openglyph-ocr-0.safetensors (the arrays trained on top of that
model), tokens.json, openglyph_ocr.py, modeling_openglyph_ocr.py,
configuration_openglyph_ocr.py, requirements.txt, release.json and README.md -- are the
work of the OpenGlyph project. That work, and the changes it made in the derived
files, are licensed under the Apache License, Version 2.0 (see LICENSE).
Changes made to Xiaomi-OCR-0:
- LoRA arrays are added to every linear map of the last six language layers;
- the last four vision blocks and the patch merger are replaced by retrained ones;
- 159,052 tokens are added to the tokenizer and to the output of the model, each with
an output row and a bias, so that the output layer is no longer tied to the input
embedding.
In base/ none of this is applied: openglyph_ocr.py applies it when it builds the model.
In model.safetensors it is applied: the LoRA arrays are folded into the 46 weights they
belong to, and those weights, the retrained vision arrays and the output layer are
stored as 32-bit floats; the input embedding has one row more for every added token
(the first of them take the place of 243 unused rows at its end); every other array is
copied from Xiaomi-OCR-0 unchanged.
Glyph names of the form <gw:NAME> are names of glyphs in GlyphWiki (https://glyphwiki.org).
This is not a product of Xiaomi or of the authors of Xiaomi-OCR-0, and is not endorsed
by them.