A curated, clustered video clip dataset built to train a from-scratch, small-parameter (~150-200M) video generation model as part of the Codeically "smallest video gen model" video series. Raw footage (batch01/ … batchNN/) — source videos as collected, each batch with its own manifest.csv: Processed clips (clips/ /.mp4 + clipsmanifest.csv) — the actual training data. Each raw source video is cut into fixed-length windows, resized, and resampled: - 2 seconds per clip, no overlap between windows - ~160px wide (height auto-scaled), landing in a 128–192px band - 10fps (20 frames per clip), landing in an 8–12fps band clipsmanifest.csv columns: clipfilename, cluster, category, source, license…