OpenNER

Small, on-device NER & de-identification models for regulated industries — banking · insurance · logistics. Fine-tuned encoders that match or beat frontier LLMs on structured, convention-bound extraction, while running on a CPU, a single GPU, Apple Silicon, or in the browser/edge via ONNX — with zero data egress.

Built and maintained by FlowX.AI. Apache-2.0. Pretrained models: huggingface.co/flowxai.

Install

pip install flowx-openner            # core inference
pip install "flowx-openner[serve]"   # + FastAPI server
pip install "flowx-openner[all]"     # + training, ONNX export, benchmarking

Quickstart

Load any published model straight from the Hub by its repo id, or a local directory:

from openner import analyze

res = analyze("Wire EUR 12,500 to DE44 5001 0517 5407 3249 31", model="flowxai/ibandetect")
print(res.entities)                       # [{'type': 'IBAN', 'text': 'DE44 ...', 'valid': True, ...}]

# De-identify with a compliance profile → redacted text + a signed audit report
deid = res.deidentify("pci-glba")
print(deid.text)                          # "Wire EUR 12,500 to [IBAN]"
print(deid.report.to_json())

Detected identifiers are validated by checksum — IBAN mod-97, card Luhn, ISIN/LEI (ISO-7064), container (ISO-6346), VIN (ISO-3779), and national IDs (CNP, EGN, OIB, EMŠO, Steuer-ID, NIR, Codice Fiscale, HU tax id) — a correctness guarantee general LLMs lack.

De-identification fails safe: every detected candidate is redacted even if its checksum fails; the valid flag is recorded in the audit report but never suppresses redaction.

Models

33 models grouped into five collections on the Hub. Every model is a fine-tuned encoder (token- or text-classification, plus one LayoutLMv3 Document-AI model), Apache-2.0, with its own model card. Load any of them by short name (analyze(text, model="mortgagedocner")) and it is pulled from flowxai/<name> on first use.

piiguard covers en, ro, bg, hu, sl, hr, de, it, fr with country-matched national-ID checksums. docformner is a LayoutLMv3 Document-AI preview trained on synthetic rendered forms; fine-tune it on real OCR'd scans before production. hscodeclassify, intentrouter and privacyfilter are text classifiers; the rest are token-classification NER.

Serve it (local REST API)

openner-serve                         # http://127.0.0.1:8000  (API + web UI)

curl localhost:8000/models
curl -X POST localhost:8000/deidentify -H 'content-type: application/json' \
  -d '{"text":"Wire to DE44 5001 0517 5407 3249 31","model":"ibandetect","profile":"pci-glba"}'

Endpoints: GET /health, GET /models, POST /analyze, POST /deidentify (redacted text + a signed audit report), POST /filter (PrivacyFilter guard: NONE / PERSONAL / FINANCIAL / HEALTH). Everything runs on-device.

Train your own

Each model is one YAML config. Train on a laptop (small slice), a CUDA box, or a cloud GPU:

python -m openner.train --config configs/banking/filingtag.yaml --profile full

Add a model = add configs/<industry>/<name>.yaml (+ a dataset adapter if new). No other code changes. Training data is synthetic and checksum-validated by construction (plus real SEC filings for filingtag).

Export for edge / phone / browser

python -m openner.export.to_onnx --model artifacts/filingtag-full
# → model.onnx (fp32) + model.int8.onnx (INT8, ~4× smaller), parity-checked against PyTorch.
# Load with onnxruntime / onnxruntime-web / onnxruntime-mobile.

Layout

openner/    industry-agnostic library (data · train · metrics · serve · deid · export · benchmark)
configs/    one YAML per model  (banking/ · insurance/ · logistics/ · cross/)
scripts/    smoke_test.sh (laptop) · train_gcp.sh (GPU) · run_benchmark.py

License

Apache-2.0. Copyright 2026 FlowX.AI. See LICENSE and NOTICE. Models are trained on synthetic, checksum-validated data; validate on your own real documents before production use.