
Extract PII character spans and confidence scores with the Fastino GLiNER2 privacy filter.
fastino/gliner2-pii detects personally identifiable information (PII) in text across 42 entity types and seven languages. Fastino developed this GLiNER2 fine-tune for identifying sensitive spans that applications can mask or redact.
This model is ready for commercial or non-commercial use.
This model is not owned or developed by NVIDIA. This model has been developed and built to a third-party's requirements for this application and use case; see link to Non-NVIDIA fastino/gliner2-pii Model Card
GOVERNING TERMS: Use of this trial service is governed by the NVIDIA API Trial Terms of Service. The gliner2-privacy-filter-PII-multi model is available under the Apache-2.0 License.
Global
Use Case: PII detection and masking, data governance, and training-data hygiene in multilingual text pipelines.
Build.NVIDIA.com: 11/16/2026 via link (planned)
Huggingface: 05/10/2026 via link
References:
Architecture Type: Transformer
Network Architecture: GLiNER2 (mDeBERTa-v3-base encoder)
Total Parameters: 307.1M
Base Model: GLiNER2
Input Types: Text
Input Formats: String
Input Parameters: One-Dimensional (1D)
Other Input Properties: Entity labels condition extraction; a subset of the 42 supported labels can be supplied.
Output Types: Text
Output Format: Other (Structured entity results)
Output Parameters: One-Dimensional (1D)
Other Output Properties: Entity results can include confidence scores and spans; application post-processing performs redaction.
Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.
Runtime Engines:
For the planned production API, send a POST request to the NVIDIA API endpoint with "model": "fastino/gliner2-pii" in the request body.
Supported Hardware:
Preferred Operating Systems: Linux
The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.
fastino/gliner2-pii
Data Modality: Text
Text Training Data Size: Undisclosed
Training Data Collection: Synthetic
Training Labeling: Synthetic
Training Properties: Fine-tuning used 4,910 synthetic texts with 129,951 PII mentions, generated with GPT-5.4 (Generative Pre-trained Transformer) at temperature 0.01 using constraint-driven generation. Languages: English, French, Spanish, German, Italian, Portuguese, and Dutch; total training tokens are undisclosed.
Testing Data Collection: Undisclosed
Testing Labeling: Undisclosed
Testing Properties: Undisclosed; the vendor does not separately describe an internal testing set's sample count, modalities, content, or language coverage.
Evaluation Data Collection: Synthetic
Evaluation Labeling: Synthetic
Evaluation Properties: The SPY: Enhancing Privacy with Synthetic PII Detection Dataset benchmark contains synthetic English text, as identified in its dataset card, representing medical consultations and legal questions; the full released corpus has 4,491 medical and 4,197 legal examples (8,688 total), generated using Llama-3-70B and the Faker library. Fastino reports exact-match span-level metrics under the alias fastino/gliner2-pii-v1; the exact sample count and entity-generation seed used for the scores below are undisclosed.
Evaluation Benchmark Score: Vendor-reported average F1 score (the harmonic mean of precision and recall): 0.477.
| Model | Legal precision | Legal recall | Legal F1 | Medical precision | Medical recall | Medical F1 | Average F1 |
|---|---|---|---|---|---|---|---|
| fastino/gliner2-pii-v1 | .346 | .750 | .473 | .369 | .686 | .480 | .477 |
| nvidia/gliner-PII | .343 | .452 | .390 | .368 | .465 | .411 | .400 |
| urchade/gliner_multi_pii-v1 | .467 | .317 | .377 | .518 | .351 | .419 | .398 |
| openai/privacy-filter | .242 | .656 | .354 | .287 | .692 | .406 | .380 |
Acceleration Engine: Other (PyTorch + GLiNER2; NVIDIA Dynamo Runtime 1.5.1 for request routing; custom FastAPI/Uvicorn gateway)
Test Hardware: NVIDIA Lovelace (L40S); 1 GPU
The vendor reports over-prediction of names, including confusion with common nouns, organizations, and products. Evaluation covers one benchmark and two domains; broader multilingual evaluation is ongoing. Synthetic training annotations have no human validation, and non-European locales and scripts remain unmeasured. The vendor recommends label-specific thresholds, dictionary filters, and domain calibration.
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here