What’s new here

This repo distributes a single compiled neural program as an artifact identified by a fixed ID (73a0e38b8bbe3427cd1d). You load it by ID, it downloads once, and all subsequent inference runs locally. You can also recompile from the plain-English spec in spec.txt if you want to adjust the behavior, provided you have a PAW account and API key.

What it does

Call paw.function("73a0e38b8bbe3427cd1d") with a text string. The program returns a JSON array of [text, type] pairs where each text is copied verbatim from the input. The nine type labels are: private_person, private_email, private_phone, private_address, private_url, private_date, account_number, secret, and other_pii.

The README reports benchmark results on the AI4Privacy dataset: 0.8464 typed-character F1, 0.9085 label-agnostic extraction F1, and 93.9% type accuracy on overlapping characters. The evaluation used three disjoint document splits to select output format and then validate the final program.

Who it’s for

Python developers who need to strip or label PII before storing or forwarding text, and who want inference to stay on the local machine. The single-function API means minimal integration work.

Try it

pip install programasweights --extra-index-url https://pypi.programasweights.com/simple/

import json
import programasweights as paw
 
detect_pii = paw.function("73a0e38b8bbe3427cd1d")
 
text = "Name: Ada Lovelace\nEmail: ada@example.com\nPIN: 4821"
print(json.loads(detect_pii(text)))

A live demo is available at https://programasweights.com/pii.

How mature is it

Created and last pushed September 2026. 35 stars, 2 forks, 1 contributor, 4 commits in the last 90 days. No releases tagged yet. MIT license. The benchmark scripts and full result logs are included in the repo under scripts/ and RESULTS.md.