kuvvi/video-ssl-nextframe-50k-v1

Video SSL next-frame reasoning data, 50,000 clips (v1)

A deterministic 50,000-clip slice of a larger self-supervised video next-frame corpus (505,632 clips). Each row is one clip: one input frame, 2 to 6 target frames that a model should produce, and interleaved reasoning text written between the frames. Frames are 608 px wide JPEGs.

Contents

1. video_share_50k_Kuvvius_v1.part000.tar to part021.tar: 22 independent tar parts of about 400 MB (10.42 GB in total). Extract every part into the same directory to rebuild the folder video_share_50k_Kuvvius_v1/ with training/training.jsonl and training/images/<shard>/<frame>.jpg. parts.jsonl lists each part with its sha256, size and the range of files it holds.
2. training/training.jsonl (also inside part000): 50,000 rows, one JSON object per line, image paths relative to training/. Fields: id, source, task, question, answer, images (the input frame), reasoning_thought_k and reasoning_image_k (the interleaved text and target frames, k from 0), metadata (shard, n_frames, frame_paths, generation settings).
3. sampling_record.json: the selection rule, seed, per-source counts and sha256 digests. files.txt: the 204,518 frame paths. MANIFEST.md: counts and digests for checking a copy.

How the rows were chosen

Rows are the 50,000 whose key sha256("video_share_50k_Kuvvius_v1" + "\0" + "20261002" + "\0" + id) is smallest over the whole corpus. The rule depends only on the row ids, so the pick is reproducible from an id list in any order. Clips kept / in the corpus, by source: pandas70m 30,588 / 309,760; howto100m 15,082 / 152,314; Sekai 4,097 / 41,101; webvid10m 159 / 1,671; PEVideo 74 / 786.

Provenance and license

The frames derive from five public video sources: Panda-70M, HowTo100M, Sekai, WebVid-10M and PE-Video. Each has its own license and terms; several are research-only or non-commercial. This release is for research use under those upstream terms (license: other). Access is approved by hand. The reasoning text is model-generated.

Reading the rows

Each row can be turned into a conversation: the user turn holds question and the input frame in images; the assistant turn alternates reasoning_thought_0, reasoning_image_0, reasoning_thought_1, ... Image paths resolve against the training/ directory.