All learning paths Learn See Try Make Read a table without inventing a cell What this lesson is about Copy out a printed table and a filled form exactly, and mark every box you genuinely cannot read.
A scanned land record: a printed grid, a filled form, and one box hidden under a stamp. How do you type it?
A Type every box that holds something, and skip the blank ones. B Type the printed grid first and the filled form after, since one is easier. C Copy a table exactly, empty cells and reading order included.
Check my recall Why this matters The PubTables work names the common fault: the same table typed out two different ways. One written rule, used on every row, is what makes two people's work fit together.
Type out a printed table and a filled form without inventing a single cell.
The PubTables work names the common fault: the same table typed out two different ways. One written rule, used on every row, is what makes two people's work fit together.
Copy a table exactly, empty cells and reading order included. Handwriting, stamps and overlapping text break document work. Mark them, do not guess. Report the share of the page you could not read; that number is the batch quality. Task A scanned land record: a printed table on top, a handwritten form below, one cell under a stamp.
Weak approach Typing the likely number under the stamp, and leaving out two empty cells to save time.
Stronger approach Every cell typed in reading order, empty ones included. The cell under the stamp is marked unreadable. The note says four of sixty cells could not be read.
Why the stronger approach works A guessed number in a land record is worse than a blank one. Nobody later can see that it was a guess.
Try a changed situation The page is now a handwritten attendance register photographed on a phone. Two names are written on top of an older entry that was rubbed out. What do you do?
Make something yourself Type out one printed table and one filled form exactly as they stand. Mark every box you could not read, and count them as a share of the whole page.
Lesson 7 of 10 on Multimodal & physical AI next Judge a generated image against its prompt Part 2 ends in a work sample you can send The rules this work sample is read against
Read the primary or official source: Smock et al., PubTables-1M (arXiv)