v0.5 — W8A8 FP8 (compressed-tensors, calibrated on 480 real pages), v0.5: v0.4.1 plus two supervised fine-tuning stages and a one-direction repetition-loop edit (1,024 MLP output columns, 0.066 % of weights); OmniDocBench v1.6 overall 77.12 to 83.45; requires shrew-server >= 0.5.0
13508bc verified Download processor_config.json from btbtyler09/shrew-ocr-preview-FP8: direct link, hf CLI and curl.
- Browser
- Download file 286 Bytes
-
https://huggingface.co/btbtyler09/shrew-ocr-preview-FP8/resolve/main/processor_config.json
- Command line
-
hf download hf://btbtyler09/shrew-ocr-preview-FP8/processor_config.json
-
curl -L -o processor_config.json https://huggingface.co/btbtyler09/shrew-ocr-preview-FP8/resolve/main/processor_config.json
286 Bytes
| { | |
| "downsample_rate": "4/8", | |
| "image_token": "<image>", | |
| "num_additional_image_tokens": 0, | |
| "patch_size": 16, | |
| "processor_class": "Granite4VisionProcessor", | |
| "auto_map": { | |
| "AutoProcessor": "processing.Granite4VisionProcessor" | |
| }, | |
| "vision_feature_select_strategy": "full" | |
| } | |