Limits and evaluation¶
Use this page to decide what the published measurement does and does not say. The figures below are the ones on the listing.
What was measured¶
On a held-out split of 648 Portuguese sentences, the listing says the model removed 98% of 915 labelled items, made 19 marks that were not in those labels, and wrongly marked 6 of 80 sentences that had no personal data or only look-alike numbers. The sentences are described as public court decisions, Wikipedia-style text, and support, banking, health, legal, and vehicle messages that used fake but valid identifiers.
RG, RENAVAM, and vehicle plates were each found in full on a second held-out split.
On ml.c7i.xlarge, the listing says one short message takes about 0.3
seconds and one instance cleans about 31,000 messages an hour when every
worker is busy. That is a measurement of short messages with every worker
busy, not a throughput commitment.
What that measurement is not¶
It is not a promise that every personal-data item in your text will be removed. It is not a certification of LGPD compliance, and it is not a clinical or healthcare evaluation. The health messages in the test set are described as fake identifiers, not as a clinical study.
Names in all-lowercase informal text are found less reliably than in normal text. RG numbers and bank details have no national check digit and are recognised by the words around them. Order, protocol, and court-case numbers are a stated reason the rules look at check digits and the words around a number. Look-alike numbers can still be marked, as the 6-of-80 result shows.
Placeholder replacement is pseudonymisation. Anyone who has the key can relate the placeholder back to the class of value the model saw. Keep the key secret. This is not anonymisation.
The model accepts Portuguese text. Other languages are outside the published scope.
Input bounds¶
A real-time request should stay near 50 short texts or 20,000 characters,
inside SageMaker's 60-second limit. A batch record can be up to 1 MB.
Original values appear in the response only when include_values is true.
Leave it false unless you have a place to store those values.
Last updated November 11, 2026 · Suggest an improvement