Evaluate your reader on the corpus

Real financial documents, a truth no reader produced, and a score that counts the readings your reader gave as reliable and got wrong. Under licence, not for sale yet.

What you receive

• 50 documents in PDF, one per issuer: financial tables, and report pages whose words sit at fixed positions with no table underneath. Half from French issuers, half from the rest of the European Economic Area.
• The truth of each one, taken from the issuer’s own iXBRL tagging of the published report: the rows of numbers, in printed order.
• The scoring script, one Node command with no dependency, and the same judge truecopy itself is measured with.
• The readings of two public readers and their scores, so you can check that your setup reproduces them.

What the score counts

• Documents and rows read exactly, every number in its row.
• Readings given as reliable that were wrong, the error nobody checks in production.
• The share of documents your reader can keep while staying under 5% error.
• The gap between the confidence it states and the accuracy it reaches.

The terms

• Evaluation only. Training or tuning a reader on these documents is not part of the licence unless it is added as an option.
• No redistribution: the documents and their truths stay with the signatory.
• Only documents whose publisher’s terms allow it are handed over.
• No price yet. Join the waiting list to hear when the licence opens.

Join the waiting list

You are on the list. You will get one message when the licence opens.

That address could not be read, and nothing was kept. Check it and send the form again.

One message when the licence opens, then the address is deleted. To leave earlier, write to [email protected]. What is kept.