Training

Fine-tuning multimodal models on industrial data

What changes when you need a strict condition check on an image rather than a generated description.

·12 min read

Multimodal models describe a scene well in free form, but the production task is usually phrased differently: is this specific condition met in this frame. That is closer to classification than to generation.

The first fix is to constrain the answer with a schema and validate the output. Free text turns quality assessment into a separate research problem, while a strict schema lets you compute familiar metrics.

The second is to collect negative examples carefully. A model trained only on frames containing a violation confidently finds violations where there are none.

Adapter-based fine-tuning is usually enough: a full fine-tune rarely pays off on a sample of a few thousand examples and makes rolling back to the previous version noticeably harder.

All articles