A completed page is written server-side as its own small parquet shard
(data/train-OFFSET-OFFSET+99.parquet) and pushed to
— — this language's own repo, so a review pass never
touches the original source dataset. This needs the companion backend in
review_server/ running somewhere (see its README) — paste its URL below,
or open this console from the backend's own /review_console.html and it's
filled in automatically. Leave it blank to just use the manual export at the bottom instead.
Works without any server: export your edits/deletes as a JSON manifest, then run the matching Python script locally with your own write-scoped HF token.