1.1.0

The aparte-privacy v1.1 model (int8, 196 KB) bundled in ApartePrivacy: same size, same speed, same API. It fixes the known limit of 1.0, short messages: on 1,067 short chat messages, strict span F1 0.94 overall and 0.85 on names (v1: 0.73 and 0.35). On our test set: 0.92 overall and 0.77 on names (v1: 0.88 and 0.66). Same spans as the Python reader on the 200 reference texts, on macOS, Linux and the iPhone simulator; builds for watchOS, tvOS and visionOS.

Limits: names stay the weakest category; a product or service name next to a key (« ma clé Slack : … ») can be taken for a name. Numbers and details: https://huggingface.co/apartejs/aparte-privacy

1.0.0

First release: the Swift runtime of aparte-privacy, with the v1 model (int8, 196 KB) bundled: detect, mask / unmask (values), minScore, ignore, ignorePatterns. Same spans as the Python reader on the 200 reference texts, on macOS, Linux, Windows and the iPhone simulator.

Known limit: short messages. On a message of a few words it misses most names (found 9 times out of 100 in our short-message probe) and some phone numbers; inside a longer request it does far better. Names are the weakest category overall (0.66 F1); the other categories are at 0.90 and above.