I believe the pipeline_tag must be `image-text-to-text` which is the underlying task of the model in transformers/vllm/etc. The Image-to-Text label itself it not wrong but takes precedence as "main pipeline tag" and therefore breaks a few integrations with the Hub (no "conversational" tag, broken inference widget, no chat template parser, etc.)
This PR should solve this by letting the Hub auto detect the correct pipeline tag.
This should also solve this issue reported here: https://github.com/huggingface/hub-docs/issues/2481
