CompressingVLM/glm-ocr-boundingdocs-ft-bnb-nf4

Source model: /content/dce_checkpoints/gaycor/base_glm_ocr_unquantized

Quantization: bitsandbytes 4-bit.

Model kind: gaycor_tit GLM-OCR LoRA fine-tuned checkpoint-8000

This repository stores a PEFT adapter plus quantization metadata. Load the base model with the included BitsAndBytesConfig and then apply the adapter.

The accompanying notebooks evaluate with the same branch metrics used before quantization: ANLS, spatial precision/recall/F1, ANLS * Spatial F1, and bbox coverage. For the Qwen distilled model they also report value exact/contains match.