Update preprocessor_config.json
#54
by
nm-research
- opened
Gemma3ImageProcessor supports do_convert_rgb, and its docstring states the default should be True but this is not enforced in the code. And, the released preprocessor_config.json sets "do_convert_rgb": null, which results in erroneous behavior.
Testing with EleutherAI/lm-evaluation-harness - during vision evaluation the above leads to errors with RGBA images (commonly seen in ChartQA, MMMU, etc.), as no conversion is applied unless explicitly forced.
This PR sets "do_convert_rgb": true explicitly in the Gemma preprocessor config to match expected behavior and prevent errors.
Note:
I am applying this change to one of the Gemma-3 model preprocessor configs first. If accepted, I will propagate this fix to other Gemma-3 models as well.