temsa/eurocivic-gliner2-onnx-cpu-v2

EuroCivic GLiNER2 ONNX CPU

EuroCivic GLiNER2 ONNX CPU is a CPU-oriented GLiNER2 bundle for zero-shot entity extraction and label-set classification in civic and public-service text. It is based on fastino/gliner2-base-v1, then fine-tuned for civic-service intent classification, sentiment-style feedback classification, and PII extraction.

The primary language targets are English and Irish. Additional European-language coverage was kept in the training mix for German, Spanish, French, Polish, Ukrainian, Dutch, Portuguese, Italian, and Swedish. Quality is not expected to be equal across all languages; validate thresholds for your own domain.

Precision variants

The gliner2_config.json file exposes three precision keys:

Precision key	Encoder	Classifier	Span representation	Count embedding	Notes
fp32	fp32	fp32	fp32	fp32	Highest-fidelity export.
qint8	fp32	QInt8	QInt8	fp32 passthrough	Recommended CPU setting from validation. Partial q8, not full-model q8.
dynamic_q8	fp32	fp32	QInt8	fp32 passthrough	Alternative partial dynamic-q8 setting.

The quantized variants are intentionally partial. Full encoder quantization was not retained because validation quality dropped sharply. Use qint8 first unless your runtime has a specific reason to prefer dynamic_q8.

What it is tuned for

• Zero-shot entity extraction with arbitrary labels.
• PII-like entity extraction for names, email addresses, phone numbers, postal addresses, passport-like identifiers, national identifiers, and IBAN-style values.
• Civic and public-service query classification, including task/service intent, publication/report intent, update/news intent, initiative/hub intent, specialist workflow intent, recurring publication series, annual programme cycles, generic topics, and people/office/organisation queries.
• Sentiment-style feedback classification with labels such as satisfied, frustrated, and neutral.
• Irish-centric examples, including Irish-language text, Irish names with and without diacritics, and PPS/PSP-style identifiers.
• European-friendly multilingual classification and PII extraction examples.

Training data transparency

Public Hugging Face datasets used in the adapter lineage:

• nvidia/Nemotron-PII (cc-by-4.0)
• DataikuNLP/kiji-pii-training-data (apache-2.0)
• ai4privacy/pii-masking-300k (other; review the dataset card before redistribution-sensitive use)

The final tuning stage used 133,901 training rows and 5,579 validation rows. It mixed synthetic civic-intent, synthetic multilingual PII, synthetic sentiment, and local replay-derived examples. The non-public training rows are not redistributed in this repository.

Explicit language-tagged rows in the final stage included: English 72,080, Irish 8,781, German 2,989, Spanish 2,809, French 2,742, Polish 2,541, Ukrainian 2,438, and 845 each for Dutch, Portuguese, Italian, and Swedish.

See training_summary.json for the compact provenance and validation summary.

Validation snapshot

Small internal validation checks for the recommended qint8 key:

• Civic intent contrastive set: 81/135
• Civic intent smoke set: 7/9
• Sentiment smoke set: 5/6
• PII label-level smoke: precision 0.6875, recall 0.9167

These are not broad public benchmarks. Treat them as smoke/regression checks for this export, not a guarantee of production accuracy.

Files

• gliner2_config.json: precision map and export metadata.
• onnx/encoder.onnx: shared fp32 encoder used by all active precision keys.
• onnx/classifier.onnx, onnx/classifier.qint8.onnx: classification heads.
• onnx/span_rep.onnx, onnx/span_rep.qint8.onnx, onnx/span_rep.dynamic_q8.onnx: span representation heads.
• onnx/count_embed.onnx, onnx/count_embed.qint8.onnx, onnx/count_embed.dynamic_q8.onnx: count embedding component; q8-named variants are fp32 passthrough copies for precision-key completeness.
• Tokenizer and base config files needed by GLiNER2-compatible ONNX runtimes.

Runtime notes

A runtime should select files from gliner2_config.json -> onnx_files -> <precision key>. For CPU use, start with:

MODEL_ID=temsa/eurocivic-gliner2-onnx-cpu
MODEL_PRECISION=qint8

For PII extraction, tune score thresholds against your target text. For classification, use short label descriptions rather than opaque IDs where possible.

Limitations

• This is a civic/public-service specialist, not a general-domain NER benchmark leader.
• Quantized precision keys are partial because the encoder remains fp32.
• PII extraction is not a compliance guarantee and should be paired with tests, thresholds, and review for high-risk uses.
• English and Irish were prioritized; other European languages are included but less heavily weighted.