index/instructions/<split>.parquet keep the instruction content only (the transition, the click, the primary instruction, its variants and the referring expression); the manifest is removed. The transitions_preview metadata drops the same fields, with the same rows and screenshots. The card, docs/schema.md and docs/splits.md describe the data without build history.
Add natural-language instructions for recorded clicks
index/instructions/<split>.parquet: 663,635 transitions with instructions (551,651 train; 105,703 across the four test splits), each variant with its round-trip result. manifest.json records counts and hashes. The card and docs/schema.md document the table; checksums/sha256sums.txt is updated.
Add the transitions_preview subset for browsing recorded clicks
A second Dataset Viewer subset: 100 recorded clicks per split, each from a different episode, with its instruction, the screens before and after the click (byte-identical copies of the corpus PNG members), a crop of the clicked element, the action and the transition metadata. Stored as demo/transitions/<split>.tar; the card, docs/schema.md and checksums/sha256sums.txt are updated. The default subset and all existing files are unchanged.
Open each split with diverse observations; drop the known-issues note
The first shard of every split now starts with 40 observations from 40 different episodes, rotating appearance presets, resolutions, application sets and wallpapers, so the viewer preview is representative. Only the order of observations changed: every member keeps its bytes and hash. index/shard_members.parquet (byte offsets), checksums/shard_stats.parquet and checksums/sha256sums.txt are updated accordingly. docs/known_issues.md is removed.
Correct the checksum manifest and the preview config
