SurgVLMdata-new-instruct-only

This repo contains ONLY the images referenced by the 5 new tasks' train (instruct) split that are NOT already covered by SurgVLMdata-new-reasoning.

new-reasoning's image set is a 100% subset of the full train image set (356,717 of 367,654 images) since CoT rows for these tasks are sampled from the same underlying video frames as train. To avoid re-downloading that 97% overlap twice, this repo ships only the remaining delta: 10937 images.

To reconstruct the FULL train image set for these 5 tasks, combine this repo's parts/ with SurgVLMdata-new-reasoning's parts/.

Tasks covered: Adverse Event Localization, Camera Movement Prediction, Camera Movement Detection, Instrument Movement Prediction, Instrument Movement Detection.