ActivityNet

100 distinct train videos · 20 videos per caption-duration bin · 1 FPS nearest-PTS samples · GPT-6.1-Sol · no masks

Sampling: The 1 FPS schedule starts at full-video t0 and targets integer seconds. Each displayed image is the nearest decoded original frame by its timestamp; timestamps are not rounded and frame indices are not calculated by multiplying by an assumed FPS. GPT receives only frames whose actual PTS falls inside that exact original caption interval [start, end). Action support uses global 1 FPS frame indices, and exported action end seconds are clipped to the original caption's right boundary.

Caption policy: Original ActivityNet sentence text and source seconds are preserved. Each video is grouped by a qualifying source caption in the page's duration bin; that qualifying caption is marked, while all source captions for the video are shown. Every sentence longer than 1 second is sent separately to GPT with its own in-window frames. Captions of 1 second or less remain visible as source-only rows and are excluded from model annotation. No masks, boxes, or tracks are supplied or inferred. These are model drafts for human review.

Selection note: videos were sampled without replacement from the first 1.46 GB of ActivityNet's first tar shard after scanning its member headers; this is a caption-duration-stratified preview with archive-order/availability bias, not a uniform random sample of the complete train split. Original caption times are source seconds. Frame/action ranges are zero-based half-open sample indices.