Claimentine/vans-injection-data

VANS injection-experiment data backup

Backup of intermediate artifacts for a ThinkJEPA followup (JEPA world-model latent injected into a VLM), built on top of VANS-Data-100K's COIN/YouCook2 "predict what happens next" QA task. See qa_split_full.json for the train/val/test split (12,032 matched QA pairs out of the full 28,226 CSV rows -- limited by how many source YouTube videos were still downloadable).

• qa_split_full.json: QA pairs (question, correct_caption, distractor_caption)
  • relative clip references, train/val/test split.
• clips/: extracted (in, out) video segments per COIN/YouCook2 annotation, {video_id}/{annotation_id}.mp4, ffmpeg stream-copy from the raw downloads.
• vjepa_cache_full/: V-JEPA2 ViT-L (frozen, Meta's pretrained checkpoint) patch-token features per clip, (64, 128, 1024) fp16, keyed by a flattened clip path ({video_id}__{annotation_id}.npz).
• vlm_guidance_cache_full/: Qwen3-VL-2B-Thinking guidance features (vlm_old/vlm_new, 8 decoder layers) for each QA item's input clip.