Viviannnyang/eventlens-1fps-batch-frames

EventLens 1 FPS action-annotation image assets

This public dataset repository hosts only the 1 FPS JPEG frames needed as HTTPS image inputs for an OpenAI Responses Batch run. It does not contain the source MP4 files.

Source and license

The frames are derived from the public ET-Instruct-164K dataset, specifically the EventLens hier_v1 annotations in Lostgreen/EventLens-Train. The source dataset card identifies the license as CC-BY-NC-SA-4.0. These derived frames are provided under the same license; please follow its attribution, noncommercial-use, and share-alike terms.

Sampling and inclusion

• The batch preparation contains 1,053 canonical hier_v1 videos and 11,763 original captions after the duration filters below. Legacy-only records that have no original caption field are excluded.
• A video is included when its original duration is in [60,120) seconds and it has at least one original L3 caption whose duration is strictly greater than 1 second.
• The first decoded video-frame PTS defines original-video t0. For each integer-second target, the selected source frame is the first decoded frame whose exact rational relative PTS is greater than or equal to that target. No nominal-FPS frame-number conversion or rounding is used.
• Each request uses only sampled frames whose exact relative PTS is within the associated caption's original L3 interval [start,end).

Attribution

Please cite and credit the upstream ET-Instruct-164K and EventLens dataset authors when reusing these frames. The frame filenames retain the source video ID and global 1 FPS sample-frame index.

• Eligible videos: 1053 from canonical hier_v1, with source duration [60,120) seconds and at least one original L3 caption whose duration is strictly greater than 1 second. The full hier_v1 duration bucket contains 1,057 video records; 4 have no L3 caption left after this filter.
• Eligible L3 captions: 11763; original caption text is copied verbatim. Captions of duration <=1s are excluded.
• The previously reported 1,214 included 161 legacy-only videos with sub_action summaries and no original caption; those are listed in excluded_legacy_only_missing_caption.json and are not silently substituted.
• Prompt action rules are copied verbatim from the final COIN prompt's generic rules and V5 clarification sections; COIN-specific caption-generation/source-label instructions are excluded. EventLens provides its own immutable-caption and PTS-bounded temporal instructions.
• No source videos, 1 FPS frames, image URLs, or API requests have yet been prepared/submitted.