data/ThinkMorph/README.md

ThinkMorph local data declarations

The released Parquet files and ThinkMorph/ UniAR cache contain all 24,990 training rows. They have not been repartitioned or overwritten.

ThinkMorph-IDVal5/ is our fixed, seed-17, approximately 95/5 development view, not an official ThinkMorph validation/test split. It references the same tensor shards without copying or re-encoding them.

Task	Development train	Validation	Full-data train
Chart Refocus	5,700	300	6,000
Jigsaw Assembly	5,700	300	6,000
Spatial Navigation	5,700	300	6,000
Visual Search	6,641	349	6,990
Total	23,741	1,249	24,990

Selection is by whole connected groups of source-image families and decoded input/target image identities, stratified by task and available subtypes (jigsaw grid, maze size, chart/search source type). Conservative near-identical natural-image re-encodings/resizes are grouped too. There are 28 multirow image groups; none crosses the split. Arbitrary crops or semantically identical rerenders are not exhaustively detected.

• ThinkMorph-IDVal5/manifest.jsonl: fixed train/validation assignments.
• ThinkMorph-IDVal5/image_groups.jsonl: source paths, pixel hashes, group IDs, and subtype provenance, independent of answers/model predictions.
• ThinkMorph-IDVal5/SPLIT.json: source hashes, policy, counts, and leakage audit.
• ThinkMorph-Grid16/: parallel all-train cache with understanding-aligned user image preprocessing and every generated target centered on a 512×512 gray canvas before fixed 16×16 BSQ encoding.
• ThinkMorph-Grid16-IDVal5/: zero-copy development view that reuses the exact original IDVal5 sample/group assignments while pointing to Grid16 tensor shards.
• LatentUM-IDVal5/: standalone zero-copy LatentUM SFT manifest containing only sample/task/split/group and raw-Parquet coordinates. It preserves the same assignments, never reads UniAR tensor shards, and has manifest SHA256 3559c6aa1ab3e7f09731c9a3161f553e76e1ba079472b050ebf4b37892bfe3e1.

Use the development view only for configuration/epoch selection. Final full-data training uses ThinkMorph/, disables internal validation, and starts again from the declared base model. Once validation rows enter final training, they are not held out and cannot support a validation generalization claim.

The owning workflow, configs, selection caveats, and reproduction command are in docs/thinkmorph.md. Existing fixed split artifacts must be reused, not regenerated with a different seed or silently overwritten.