AVTime

avtime-longvale-pct100-v1 is a restartable archival release of full-video temporal-event training data. It contains all available AVTime completed episodes and the LongVALE training split, each rewritten as one complete original video and one compact JSON answer. The same data can be used for SFT or as shared training data in RL workflows.

Label format

Each row in train.jsonl has exactly one videos entry and three messages. The assistant target is:

{"events":[{"start_pct":0,"end_pct":12,"caption":"..."}]}

start_pct and end_pct are inclusive integer buckets from 0 through 99, relative to the full original video duration. Every event has only start_pct, end_pct, and caption.

Contents

• 57,217 training rows
• 57,217 unique original videos
• 2173.31 GiB of original media before compression
• 1,294 independently extractable media archives
• 1 unusually large original video(s) stored as byte-identical parts

The release intentionally excludes historical event-window helper clips. They are redundant once each training episode is represented by one complete video.

Restore

First extract the metadata archive, then extract every media archive into the same destination:

for archive in archives/*.tar.zst; do
  zstd -t "$archive"
  tar -I zstd -xf "$archive"
done
python restore_split_media.py --root .

After extraction, train.jsonl references media with package-relative paths such as media/longvale/... and media/avtime/.... The restore command is a no-op for normal files and concatenates only the exceptional .parts videos without transcoding or changing their bytes.

Source and redistribution notice

This release is derived from AVTime and LongVALE training data. It preserves no extra license grant for original video assets. It is public at the release owner's direction; users are responsible for complying with applicable upstream redistribution terms.