Upload 5 files

BIAS.md

@@ -0,0 +1,5 @@
+Field                                                                                               |  Response
+:---------------------------------------------------------------------------------------------------|:---------------
+Participation considerations from adversely impacted groups [protected classes](https://www.senate.ca.gov/content/protected-classes) in model design and testing:  |  Not applicable for the training phase, as the model was trained exclusively on synthetic data. Stakeholder review with impacted groups is recommended during deployment to validate real-world performance.
+Bias Metric (If Measured):  |  Strict F1 Score at a 0.3 threshold. Performance on evaluation datasets was: Argilla PII (0.70), AI4Privacy (0.64), and nvidia/Nemotron-PII (0.87). Demographic performance breakdowns are not available.
+Measures taken to mitigate against unwanted bias: | Not applicable.

EXPLAINABILITY.md

@@ -0,0 +1,13 @@
+Field                                                                                                  |  Response
+:------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------
+Intended Task/Domain:                                                                                  |  PII/PHI Detection: To detect and classify Personally Identifiable Information (PII) and Protected Health Information (PHI) in structured and unstructured text across domains like healthcare, finance, and legal.
+Model Type:                                                                                            |  Transformer (GLiNER architecture).
+Intended Users:                                                                                        |  Developers and data professionals implementing data governance, privacy compliance (GDPR, HIPAA), and content moderation workflows.
+Output:                                                                                                |  A list of dictionaries, where each dictionary contains the detected text, its label (e.g., SSN), start and end positions, and a confidence score.
+Describe how the model works:                                                                          |  The model takes a text string as input and uses a non-generative transformer architecture to produce span-level entity annotations. It identifies and labels sensitive information across 55+ categories without generating new text.
+Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of:  |  Not Applicable
+Technical Limitations & Mitigation:                                                                    |  Limitation: Performance varies by domain, text format, and the confidence threshold chosen. Mitigation: NVIDIA recommends use-case-specific validation and human review for high-stakes deployments to ensure accuracy and safety.   
+Verified to have met prescribed NVIDIA quality standards:  |  Yes
+Performance Metrics:                                                                                   |  Strict F1 Score is the primary evaluation metric. The model also provides per-entity confidence scores in its output.
+Potential Known Risks:                                                                                 |  If the model does not work as intended, it could lead to false negatives (failing to detect PII) or false positives (incorrectly flagging non-sensitive data, causing unnecessary redaction).
+Licensing:                                                                                             |  Use of this model is governed by the [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/)

PRIVACY.md

@@ -0,0 +1,9 @@
+Field                                                                                                                              |  Response
+:----------------------------------------------------------------------------------------------------------------------------------|:-----------------------------------------------
+Generatable or reverse engineerable personal data?                                                     |  No
+Personal data used to create this model?                                                                                       |  No
+How often is dataset reviewed?                                                                                                     |  The dataset was reviewed during its creation, model training, evaluation, and before release.
+Is there provenance for all datasets used in training?                                                                                |  Yes
+Does data labeling (annotation, metadata) comply with privacy laws?                                                                |  Yes. Labels were automatically injected during the synthetic data generation process, so no real personal data was ever viewed or handled.
+Is data compliant with data subject requests for data correction or removal, if such a request was made?                           |  Not Applicable.
+Applicable Privacy Policy        | https://www.nvidia.com/en-us/about-nvidia/privacy-policy/

README.md

@@ -1,162 +1,122 @@
+# GLiNER-PII Model Overview
+
+### Description:
+GLiNER-PII is a successor to the Gretel GLiNER PII/PHI models. Built on the GLiNER bi-large base, it detects and classifies a broad range of Personally Identifiable Information (PII) and Protected Health Information (PHI) in structured and unstructured text. It is non-generative and produces span-level entity annotations with confidence scores across 55+ categories. This model was developed by NVIDIA.
+
+This model is ready for commercial/non-commercial use. <br>
+
+### License/Terms of Use
+Use of this model is governed by the [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/). <br>
+
+### Deployment Geography:
+Global <br>
+
+### Use Case: <br>
+GLiNER-PII supports detection and redaction of sensitive information across regulated and enterprise scenarios. 
+
+- **Healthcare**: Redact PHI in clinical notes, reports, and medical documents.
+- **Finance**: Identify account numbers, SSNs, and transaction details in banking and insurance documents.
+- **Legal**: Protect client information in contracts, filings, and discovery materials.
+- **Enterprise Data Governance**: Scan documents, emails, and data stores for sensitive information.
+- **Data Privacy Compliance**: Support GDPR, HIPAA, and CCPA workflows across varied document types.
+- **Cybersecurity**: Detect sensitive data in logs, security reports, and incident records.
+- **Content Moderation**: Flag personal information in user-generated content.
+
+Note: performance varies by domain, format, and threshold, so validation and human review are recommended for high‑stakes deployments. <br>
+
+### Release Date:  <br>
+Hugging Face 10/30/2025 via https://huggingface.co/nvidia/gliner-pii <br> 
+
+## References(s):
+- GLiNER base (Hugging Face): https://huggingface.co/knowledgator/gliner-bi-large-v1.0
+- Gretel GLiNER PII/PHI models: https://huggingface.co/gretelai/gretel-gliner-bi-large-v1.0
+- Training dataset: https://huggingface.co/datasets/nvidia/nemotron-pii
+- GLiNER library: https://pypi.org/project/gliner/
+
+## Model Architecture:
+**Architecture Type:** Transformer <br>
+
+**Network Architecture:** GLiNER <br>
+
+**This model was developed based on knowledgator/gliner-bi-large-v1.0** <br> 
+**Number of model parameters: 5.7 × 10^8** <br>
+
+## Input: <br>
+**Input Type(s):** Text <br>
+**Input Format:** UTF-8 string(s) <br>
+**Input Parameters:** One-Dimensional (1D) <br>
+**Other Properties Related to Input:** supports structured and unstructured text <br>
+
+## Output: <br>
+**Output Type(s):** Text <br>
+**Output Format:** String <br>
+**Output Parameters:** One-Dimensional (1D) <br>
+**Other Properties Related to Output:** List of dictionaries with keys {text, label, start, end, score} <br>
+
+## Software Integration:
+**Runtime Engine(s):** 
+* PyTorch, GLiNER Python library <br> 
+
+**Supported Hardware Microarchitecture Compatibility:** <br>
+* NVIDIA Ampere <br>
+* NVIDIA Blackwell <br>
+* NVIDIA Hopper <br>
+* NVIDIA Lovelace <br>
+* NVIDIA Pascal <br>
+* NVIDIA Turing <br>
+* NVIDIA Volta <br>
+* CPU (x86_64) <br>
+
+**Preferred/Supported Operating System(s):**
+* Linux <br>
+
+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. <br>
+
+## Model Version(s):
+- nvidia/gliner-pii
+- Version: v1.0
+
+## Training and Evaluation Datasets:
+
+### Training Dataset
+
+**Link:** [nvidia/nemotron-pii](https://huggingface.co/datasets/nvidia/nemotron-pii) <br>
+**Data Modality:** Text <br>
+**Text Training Data Size:** \~100k records (\~10^5, <1B tokens) <br>
+**Data Collection Method:** Synthetic <br>
+**Labeling Method:** Synthetic <br>
+
+**Properties:**
+Synthetic persona-grounded dataset generated with NVIDIA NeMo Data Designer, spanning 50+ industries and 55+ entity types (U.S. and international formats). Includes both structured and unstructured records. Labels automatically injected during generation.
+
+## Evaluation Datasets
+
+* [Argilla PII](https://huggingface.co/argilla)
+* [AI4Privacy](https://huggingface.co/ai4privacy)
+* [Gretel PII Dataset V1/V2](https://huggingface.co/datasets/gretelai/gretel-pii-masking-en-v1)
+
+**Data Collection Method:** Hybrid: Automated, Human <br>
+**Labeling Method:** Hybrid: Automated, Human <br>
+
+**Evaluation Results** <br>
+From the combined evaluation across Argilla, AI4Privacy, and Gretel PII datasets:
+
+| Benchmark           | GLiNER V1 (Strict F1) | GLiNER V2 (Strict F1) |
+| --------------------| --------------------: | --------------------: |
+| Argilla PII         |                  0.64 |                  0.70 |
+| AI4Privacy          |                  0.60 |                  0.64 |
+| nvidia/Nemotron-PII |                  0.66 |                  0.87 |
 ---
-license: apache-2.0
-language:
-  - en
-library_name: gliner
-datasets:
-  - nvidia/Nemotron-PII
-pipeline_tag: token-classification
-tags:
-  - PII
-  - PHI
-  - GLiNER
-  - information extraction
-  - encoder
-  - entity recognition
-  - privacy
----
 
-# GLiNER-PII: Fine-Tuned Model for PII/PHI Detection
-
-The GLiNER-PII is a fine-tuned successor to the Gretel GLiNER PII models. Built on the GLiNER bi-large base (`knowledgator/gliner-bi-large-v1.0`), it detects and classifies a broad range of Personally Identifiable Information (PII) and Protected Health Information (PHI) in **English text**. The model works with both structured and unstructured text and is non-generative, producing span-level entity annotations with confidence scores across 55+ categories.
-
-This model is intended for privacy-preserving NLP workflows such as de-identification, redaction, and compliance checks in healthcare, finance, legal, and enterprise data pipelines.
-
-For more information about the base GLiNER model, including its architecture and general capabilities, please refer to the [GLiNER Model Card](https://huggingface.co/knowledgator/gliner-bi-large-v1.0).
-
-## Training Data
-
-The model was fine-tuned on the `nvidia/Nemotron-PII` dataset, a synthetic, persona-grounded dataset containing 100,000 records across 50+ industries with span-level annotations for 55+ PII/PHI categories. The dataset was generated with NVIDIA NeMo Data Designer using synthetic personas grounded in U.S. Census data to ensure demographic realism and contextual consistency.
-
-**Dataset Details:**
-- **Size:** 100,000 records (50k train / 50k test)
-- **Domains:** 50+ industries (healthcare, finance, cybersecurity, etc.)
-- **Entity Types:** 55+ PII/PHI categories
-- **Locale Coverage:** US and international formats
-- **Content Types:** Both structured (forms, invoices) and unstructured (emails, notes) documents
-
-For detailed statistics on the dataset, visit the [dataset documentation on Hugging Face](https://huggingface.co/datasets/nvidia/Nemotron-PII).
-
-## Use Cases
-
-The GLiNER-PII supports detection and redaction of sensitive information across regulated and enterprise scenarios:
-
-- **Healthcare**: Redact PHI in clinical notes, reports, and medical documents
-- **Finance**: Identify account numbers, SSNs, and transaction details in banking and insurance documents
-- **Legal**: Protect client information in contracts, filings, and discovery materials
-- **Enterprise Data Governance**: Scan documents, emails, and data stores for sensitive information
-- **Data Privacy Compliance**: Support GDPR, HIPAA, and CCPA workflows across varied document types
-- **Cybersecurity**: Detect sensitive data in logs, security reports, and incident records
-- **Content Moderation**: Flag personal information in user-generated content
-
-Note: performance varies by domain, format, and threshold, so validation and human review are recommended for high-stakes deployments.
-
-## Installation & Usage
-
-Ensure you have Python installed. Then, install or update the `gliner` package:
-
-```python
-import json
-from gliner import GLiNER
-
-# Load the fine-tuned GLiNER model
-model = GLiNER.from_pretrained("nvidia/gliner-pii")
-
-# Sample text containing PII/PHI entities
-text = """
-   **Claim Denial Letter**
-
-   **Date:** 2023-11-15
-
-   **Claimant Information:**
-   - **Claimant Name:** Nataly White
-   - **Medical Record Number:** 1842-75-3924
-
-   **Claim Denial Details:**
-
-   Dear Nataly White,
-
-   We are writing to inform you that your disability claim has been denied. The denial is based on the information provided in your medical record, numbered 1842-75-3924.
-   After a thorough review, it has been determined that the medical evidence does not meet the criteria outlined in your policy for disability benefits.
-"""
-
-# Define the labels for PII/PHI entities
-labels = [
-    "certificate_license_number",
-    "first_name",
-    "date_of_birth",
-    "ssn",
-    "medical_record_number",
-    "password",
-    "unique_id",
-    "phone_number",
-    "national_id",
-    "swift_bic",
-    "company_name",
-    "country",
-    "license_plate",
-    "tax_id",
-    "employee_id",
-    "pin" ,
-    "state",
-    "email",
-    "date_time",
-    "api_key",
-    "biometric_identifier",
-    "credit_debit_card",
-    "coordinate",
-    "device_identifier",
-    "city",
-    "postcode",
-    "bank_routing_number",
-    "vehicle_identifier",
-    "health_plan_beneficiary_number",
-    "url",
-    "ipv4",
-    "last_name",
-    "cvv" ,
-    "customer_id",
-    "date",
-    "user_name",
-    "street_address",
-    "ipv6",
-    "account_number",
-    "time",
-    "age",
-    "fax_number",
-    "county",
-    "gender",
-    "sexuality",
-    "political_view",
-    "race_ethnicity",
-    "religious_belief",
-    "language",
-    "blood_type",
-    "mac_address",
-    "http_cookie",
-    "employment_status",
-    "education_level",
-    "occupation"
-]
-
-# Predict entities with a confidence threshold of 0.7
-entities = model.predict_entities(text, labels, threshold=0.3)
-
-# Display the detected entities
-print(f"{'start':>5}  {'end':>5}  {'text':<16}  {'label'}")
-print("-" * 60)
-for e in entities:
-    print(f"{e['start']:>5}  {e['end']:>5}  {e['text']:<16}  {e['label']}")
-
-# Expected output:
-# start    end  text              label
-# ------------------------------------------------------------
-# 42      52   2023-11-15        date
-# 107    113   Nataly            first_name
-# 114    119   White             last_name
-# 152    164   1842-75-3924      medical_record_number
-# 204    210   Nataly            first_name
-# 211    216   White             last_name
-# 376    388   1842-75-3924      medical_record_number
-```
+We evaluated the model using `threshold=0.3`. <br>
+
+# Inference:
+**Acceleration Engine:** PyTorch (via Hugging Face Transformers) <br>
+**Test Hardware:** NVIDIA A100 (Ampere, PCIe/SXM) <br>
+
+## Ethical Considerations:
+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. When downloaded or used in accordance with our terms of service, 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. <br> 
+
+For more detailed information on ethical considerations for this model, please see the the Bias, Explainability, Safety & Security, and Privacy Subcards. <br> 
 
+Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).  <br>

SAFETY&SECURITY.md

@@ -0,0 +1,6 @@
+Field                                               |  Response
+:---------------------------------------------------|:----------------------------------
+Model Application Field(s):                               |  Healthcare, Finance, Legal, Enterprise Data Governance, Data Privacy Compliance, Cybersecurity, Content Moderation
+Describe the life critical impact (if present).   |  Not Applicable.
+Use Case Restrictions:                              |  Abide by [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/)
+Model and dataset restrictions:            |  The Principle of least privilege (PoLP) is applied limiting access for dataset generation and model development.  Restrictions enforce dataset access during training, and dataset license constraints adhered to.