EXTRACT

1,000 messy documents. Hidden labels. Highest F1 wins.

Get notified →

$1337 enter extract

Opens: Oct 5
Submissions close: Oct 31, 11:59pm ET
Judging: Nov 1 – 7
Results: Nov 8
Entries: None yet
Judged by: F1 90 · Cost 10

Open Oct 5
Submissions close Oct 31
Judging Nov 1
Results Nov 8

[01]Brief

Pull structured fields out of 1,000 real-world documents: invoices, receipts and contracts, scanned, skewed and inconsistent. Use any models or pipelines you want. Submit a predictions file; it is scored automatically against hidden labels and the leaderboard updates instantly.

[02]Rules

Dataset and schema release Oct 5.
Submit before Oct 31, 11:59pm ET.
Predictions as JSONL (one object per document id).
Score = field-level F1 against hidden labels; declared cost per document breaks ties.
Up to 5 scored submissions per day.
Hand-labelling the test set is disqualifying.

[03]Leaderboard opens Oct 5

[ Opens Oct 5 ]

Be the first to enter EXTRACT

Entries open Oct 5. Get set up now and you can submit the moment it goes live.

Get notified→$1337 enter extract