| # | Masked text | Names hidden | Missed | Masked but not private |
|---|
For a write-up
The same results as a Markdown section, so the numbers are copied rather than retyped.
How many names does it really hide?
This page runs the same masking as Paste-Safe over made-up requests of the kind people paste into a chat AI. Every person, company, place, email, phone number and account number in them is labelled. Some are meant to be hard: names in lower case, names that are also ordinary words, and names from outside the model's English news training. Everything runs in this tab; nothing is uploaded. The messages and labels are in eval-data.js.
A name counts as hidden only if every letter of every mention is masked: "[PERSON_1] Okafor" counts as a miss for "Daniel Okafor". The always-mask list is left empty here, so this measures the model, the patterns and the name rules alone. It is scored three ways: the default, the model with the name rules off, and a lower model confidence. The table shows the default.
| # | Masked text | Names hidden | Missed | Masked but not private |
|---|
The same results as a Markdown section, so the numbers are copied rather than retyped.