TRAVL (Trajectory-Aware Video-Language)

TRAVL: Trajectory-Aware Video-Language Finetuning Dataset TRAVL is a finetuning dataset of short real vs implausible videos with targeted Q/A annotations probing physical plausibility (e.g., object permanence, collisions, shadows, trajectory consistency). Repository layout Data schema (current JSON) TRAVL QA.json is a JSON list. Each item: { “id”: “video 000”, “video”: “video 000.mp4”, “conversations”: {“from”: “human”, “value”: ” nDoes the… See the full description on the dataset page:

Source: Hugging Face Hub (INSAIT-Institute/TRAVL). Metadata imported from the dataset’s Hub tags.