---
library_name: gliner
pipeline_tag: token-classification
base_model: Qwen/Qwen3-0.6B
tags:
- gliner
- gliner-streaming
- named-entity-recognition
- zero-shot-ner
- pii
- privacy
- multilingual
language:
- en
metrics:
- precision
- recall
- f1
license: apache-2.0
---

# GLiNER Streaming PII — Qwen3 0.6B

An open-label PII detector built with the GLiNER streaming-span architecture and a Qwen3-0.6B causal backbone. It supports regular full-text inference, cached incremental streams, and full-session recomputation. This model was developed in collaboration between [Wordcab](https://wordcab.com/) and [Knowledgator](https://www.knowledgator.com/). For enterprise-ready, specialized PII/PHI/PCI models, contact us at info@knowledgator.com.
## Architecture

Following the component schema in the [GLiNER architecture docs](https://github.com/urchade/GLiNER/blob/main/docs/add_custom_architecture.md), this checkpoint is composed of:

[![GLiNER Streaming Span architecture](assets/gliner-streaming-architecture.svg)](assets/gliner-streaming-architecture.svg)


## Install and load

```bash
pip install "gliner>=0.2.28"
```

```python
import torch
from gliner import GLiNER

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = "bf16" if device == "cuda" else "fp32"

model = GLiNER.from_pretrained(
    "knowledgator/gliner-stream-pii-v1.0",
    load_tokenizer=True,
    map_location=device,
    dtype=dtype,
).eval()

labels = [
    "person",
    "email address",
    "phone number",
    "street address",
    "credit card number",
    "passport number",
]
```

## Three inference strategies

| Strategy | API | What is computed | Best for |
|---|---|---|---|
| **1. Stateless full text** | No `session_id` | Complete input and label prompt | Documents, batches, independent requests |
| **2. Cached incremental** | `session_id=[id]` | New chunk only; decoder KV, labels, words, and span history are reused | Live chat, ASR, logs, token streams |
| **3. Full recompute** | `session_id=[id], recompute=True` | Accumulated session plus new chunk; cache and all spans are rebuilt | Final pass, changed labels, correction after drift |

### 1. Stateless full text

```python
entities = model.predict_entities(
    "Jane Doe can be reached at jane.doe@example.com.",
    labels,
    threshold=0.5,
)
```

### 2. Cached incremental session

```python
session_id = "call-42"

for chunk in [
    "Customer Jane",
    " Doe asked us to call",
    " +1 (415) 555-0132.",
]:
    snapshot = model.inference(
        [chunk],
        labels,
        session_id=[session_id],
        threshold=0.5,
    )[0]
    print(snapshot)
```

### 3. Full-session recompute

```python
# The next chunk must be non-empty. This reruns all accumulated text
# and also permits a changed label set.
final_labels = labels + ["account number"]
final_snapshot = model.inference(
    [" Account 12345678 was also mentioned."],
    final_labels,
    session_id=[session_id],
    recompute=True,
    threshold=0.5,
)[0]

model.clear_session(session_id)
```

Streaming details:

- Each call returns the **complete current session snapshot**, not only new entities.
- Chunks are concatenated verbatim; preserve boundary spaces and punctuation.
- Offsets refer to the complete accumulated text.
- Keep labels fixed unless using `recompute=True`; always clear finished sessions.

## Evaluation

PIIMB ranking scores are label-agnostic, character-level masking metrics. They are micro-averaged within each task; group rows are unweighted averages across tasks. F2 is primary because it weights recall more heavily. Strict NER F1 also requires matching entity boundaries and type.

| Scope / task | Precision | Recall | Masking F1 | Masking F2 | FPR | Strict NER F1 |
|---|---:|---:|---:|---:|---:|---:|
| **English average** | 87.36% | 91.55% | 89.18% | **90.53%** | 3.01% | — |
| **Multilingual average** | 53.84% | 78.32% | 60.45% | **68.21%** | 5.85% | — |
| `ai4privacy-en` | 94.69% | 95.16% | 94.92% | **95.06%** | 1.52% | 67.99% |
| `ai4privacy-multi` | 87.34% | 92.96% | 90.07% | **91.78%** | 3.82% | 55.39% |
| `gretel` | 86.95% | 95.58% | 91.06% | **93.72%** | 5.20% | 67.38% |
| `mapa-eur-lex` | 20.34% | 63.67% | 30.83% | **44.65%** | 7.88% | 10.91% |
| `nemotron-pii` | 72.75% | 88.07% | 79.68% | **84.51%** | 4.98% | 68.93% |
| `privy` | 95.07% | 87.39% | 91.07% | **88.83%** | 0.33% | 81.68% |

Configuration: PIIMB v0.3.0, dataset revision `4a13e9ffe6fd0d275efbde8afd4d8d8f1ffc2133`, `sentences` subset, threshold 0.5, bfloat16, evaluated 2026-07-24.

The main weakness is multilingual legal and administrative text: `mapa-eur-lex` reaches only 44.65% masking F2. Validate on the target languages, domains, labels, and threshold before deployment.

## References
- [Blog](https://medium.com/p/11aefa0e4ad8?postPublishedType=initial)
- [GLiNER paper](https://arxiv.org/abs/2311.08526)
- [Qwen3 Technical Report](https://arxiv.org/abs/2505.09388)
- [PIIMB benchmark](https://huggingface.co/datasets/piimb/pii-masking-benchmark)
- [GLiNER repository](https://github.com/urchade/GLiNER)