Modelpublic

lumesec/privacy-onnx

sourceHugging Faceapache-2.0updated 22h agoView on Hugging Face

0likes

Model Card

LumeSec privacy models — ONNX

ONNX exports of open detection models, made by LumeSec for SecureGrid, a local redaction app. The models are unchanged; only the format is. One folder per model.

`gliner2-pii/`

fastino/gliner2-privacy-filter-PII-multi at revision 36126f6. Licence: Apache-2.0, as the base model; credit for the model belongs to Fastino.

file	what
`encoder.onnx`	mDeBERTa-v3 encoder. `input_ids`, `attention_mask` → `last_hidden_state`
`extractor.onnx`	span/count head, schema padded to 64 fields. `text_emb`, `schema_emb_padded`, `schema_mask`, `spans_idx` → `count_logits`, `span_scores`
`tokenizer.json`	the base model's tokenizer

Weights are stored as fp16 with a Cast to fp32 that ONNX Runtime folds at load time: half the download, fp32 arithmetic, identical results to the PyTorch model (checked span for span).

Exported with the scripts from gliner2-rs (Apache-2.0) at 1bcf40e. Reproduce with tools/models/gliner2/export.sh in the SecureGrid repository.