whodatbo1/glm-ocr-bulgarian

GLM-OCR Bulgarian (merged)

zai-org/GLM-OCR with a Bulgarian-tuned LoRA merged in. Drop-in replacement for the base GLM-OCR — same processor, same Text Recognition: prompt, same output schema.

The LoRA-only version (~50 MB) is at whodatbo1/glm-ocr-bulgarian-lora.

Use (transformers)

from transformers import AutoModel, AutoProcessor
model = AutoModel.from_pretrained(
    "whodatbo1/glm-ocr-bulgarian", trust_remote_code=True, torch_dtype="bfloat16"
)
processor = AutoProcessor.from_pretrained(
    "whodatbo1/glm-ocr-bulgarian", trust_remote_code=True
)

Use (llama.cpp / llama-server)

Convert to GGUF, patch GLM-OCR's NextN/MTP layer metadata, then serve:

python convert_hf_to_gguf.py whodatbo1/glm-ocr-bulgarian \
    --outfile glm-ocr-bg-F16.gguf --outtype f16
python convert_hf_to_gguf.py whodatbo1/glm-ocr-bulgarian \
    --mmproj --outfile mmproj.gguf --outtype f16

# GLM-OCR-specific GGUF metadata patches (see legislature_differ GUIDEBOOK)
python -m gguf.scripts.gguf_set_metadata glm-ocr-bg-F16.gguf \
    glm4.block_count 16 --force
python -m gguf.scripts.gguf_set_metadata glm-ocr-bg-F16.gguf \
    glm4.nextn_predict_layers 0 --force

llama-server -m glm-ocr-bg-F16.gguf --mmproj mmproj.gguf \
    --flash-attn off -c 4096 --image-max-tokens 1024

--flash-attn off is required (llama.cpp #19721); --image-max-tokens 1024 keeps Metal compute buffers within budget for full-page A4 inputs.

Training

See the adapter card at whodatbo1/glm-ocr-bulgarian-lora for dataset, recipe, and config.

License

Source content is in the public domain (Bulgarian official gazette). This model is released under MIT.