ner-pii-indonesian-mini

NER PII Indonesian Mini

A small Indonesian PII detector: 11.1M parameters, 11.7 MB as quantized ONNX, small enough to run in a browser tab. It is distilled from ner-roberta-large-bahasa-indonesia-finetuned, which is 2,235 MB, so this is roughly 191 times smaller and about 4 times faster on CPU.

Entity types: PER, LOC, DATE_TIME, EMAIL, PHONE, GENDER, SSN (Indonesian NIK).

Why it exists

The teacher is XLM-R large. Nearly half its parameters are a 250k piece vocabulary covering 100 languages, which is wasted on a single language task, and its label set has no I- tag for EMAIL, PHONE, SSN or GENDER, so it fragments identifiers across tokens. This student fixes both: an Indonesian only vocabulary, and a complete BIO inventory with B- and I- for all seven types.

Results

Measured on 1,626 held out sentences labelled by the teacher, so these numbers describe how closely the student reproduces the teacher rather than performance on an independent benchmark.

Per type F1 for the INT8 build:

Usage

from transformers import pipeline

ner = pipeline("token-classification", model="0xcynyx/ner-pii-indonesian-mini",
               aggregation_strategy="simple")
ner("Nama saya Budi, NIK 3204 0125 0990 0001, HP 081234567890.")

In the browser with transformers.js, the quantized ONNX build is used automatically:

import { pipeline } from "@huggingface/transformers";

const ner = await pipeline("token-classification", "0xcynyx/ner-pii-indonesian-mini");
const out = await ner("Silakan hubungi Siti di [email protected]");

Because it runs client side, text never leaves the user's device, which matters more for PII than for most tasks.

Limitations, read before deploying

• Training data is mostly synthetic. 30,923 template generated sentences with large slot pools plus 2,760 Indonesian Wikipedia lines. It is strong on form shaped text and less proven on free prose. The F1 above is agreement with the teacher on a similar distribution, not an independent benchmark.
• One epoch. Training stopped after the first epoch, which already reached parity with the teacher on this distribution. More epochs may help on natural text.
• GENDER and DATE_TIME are the weak types. GENDER had only 1,280 training examples, and DATE_TIME covers many surface forms.
• Pair it with regex for structured PII. Email, phone and NIK have exact formats, and a regex pass both repairs boundaries and catches what the model misses. The near perfect scores on those types come partly from the teacher labels having been regex corrected first.
• Not compliance evidence. Validate recall on your own data and jurisdiction before relying on it. Treat it as one control among several, not a guarantee.

Training

Character span annotations from the teacher, with production post processing applied before they became labels, so the student learns corrected boundaries rather than the teacher's fragments. One epoch, batch 16, lr 5e-5, max length 160, on an Apple M4 GPU.

Pipeline and scripts: ner-app-gradio/tools/compress.

License

MIT.