Output Explorer

Every prompt in the paper, and what each model wrote back.

Answer a question about a scanned page from its OCR text. Scored by exact match and ANLS against the accepted answers.

13 of 5,330 prompts

Nearby prompts. All 5,330 DocVQA prompts

PromptDocVQA · pagegnjk0226_1.png

What is the date on the ‘form’?

OCR text of the page · 588 characters; the scanned image itself is not published
FORM 011-371 JOM 11-70 PADS 100
To Mr.
-Larry
Mc chee
[ ] For approval of payment.
[ ] For execution.
[ ] For your information and files.
[ ] For approval as to form.
[ ] For approval as to execution.
[ ] For payment.
[ ] Note and return.
[ ] Prepare form requested.
[ ] Answer giving data requested.
[ ] Return to me with data requested.
[ ] What do you recommend?
[ ] Investigate and advise.
[ ] Are you interested?
[ ] See me about this.
[ ] Please handle.
[] What do you know of this?
Remarks:
your hole?
Date
From
state Z.
Source: https://www.industrydocuments.ucsf.edu/docs/gnjk0226
System prompt · identical for every setup
Answer the question using only the OCR text from a single document page. Return only the answer, with no explanation. Preserve the answer wording from the OCR text when possible.
Expected answer
3/31/75
Models
4 of 4 columns · click a model to add or remove it

Ours

Wrong

state Z.

8 characters4 tokens

Aux 2015

Wrong

FORM 011-371 JOM 11-70 PADS 100

To Mr. -Larry Mc chee [ ] For approval of payment. [ ] For execution. [ ] For your information and files. [ ] For approval as to

161 characters64 tokens

PiT-FT 2015

Wrong

Empty response.

0 characters

ChronoGPT 2015

Wrong
[ ] For approval of payment.
[ ] For execution.
[ ] For approval as to form.
[ ] For payment.
[ ] For execution.
[ ] For approval as to form.
[ ] For payment.
[ ] For execution.
[ ] For approval as to form.
206 characters64 tokens