Dataset for VLM fine-tuning of hand bounding box detection. Combines positive samples from HaGRID (palm + fist gestures) with negative samples from COCO (no-person and person-without-visible-wrists). Contains positive samples (hands visible) and negative samples (no hands) in conversation format compatible with Unsloth/TRL vision fine-tuning.

	Dataset Composition

Category Count Source License

Positive — 1 hand 8,578 HaGRID (palm… See the full description on the dataset page: https://huggingface.co/datasets/luksamuk/hand-tracking-vlm-dataset.