Everything on this site is public and read-only. It exists so a miner can see how their model was evaluated, what it scored, and what data it was judged on. Test content is held back for 2 days after its cycle so that nobody can be scored on a corpus they have already read.

Export

Download a released day, or any filtered slice of the corpus. The file is streamed, so a whole day is a download rather than a wait, and every file carries the manifest of every day it contains — the counts and the snapshot_content_hash — so a downloaded corpus traces back to the snapshot it came from.

What is in the file

FieldMeaning
record `manifest` for a day's header row, `content` for a corpus item.
day The UTC day the item belongs to, and the snapshot it was frozen into.
kind `test` or `production`.
id The platform's stable identifier for the item.
author_hotkey The contributing miner. Empty on production rows: production content has no author.
created_at Unix seconds, UTC.
content The item as stored.
content_encoding `verbatim`, or `csv-formula-guarded` when a CSV cell was prefixed to stop a spreadsheet executing it.
snapshot_content_hash The day's content hash, repeated on every row so a row traces back to its snapshot.

Two formats, and they are not interchangeable

JSONL

One JSON object per line, content byte-exact. This is the form to use if you intend to recompute a day's snapshot_content_hash: nothing in it is altered.

CSV

A fixed rectangle for spreadsheets. A cell that begins with =, +, - or @ would be executed as a formula by a spreadsheet, and this corpus contains text written to attack whatever opens it — so those cells are prefixed with a quote and the row's content_encoding column says csv-formula-guarded instead of verbatim. A guarded cell is not byte-identical to the stored item, which is why the CSV is not the form to hash.

An export only ever contains released content. Asking for a day inside the 2-day embargo returns that day's manifest with test_content_released set false and no test rows — a short file that says why it is short, rather than an empty one that looks broken.

Production content is unattributed. The author_hotkey column is empty on every production row because production content has no author of any kind, not because one is missing: it is the negative half of the corpus, and attributing it would hand every validator a lever on scoring.