GLiNER

// Overview

GLiNER is a framework for training and running small Named Entity Recognition (NER) models that work zero-shot. Instead of being locked to a fixed label set, you pass the entity types you want at inference time (for example "person", "award", "date"), and the model extracts matching spans from the text. It also supports joint entity and relation extraction.

It is built for teams that need entity extraction without the cost of a large language model. The models are tuned to run on CPUs and consumer hardware, and the project reports performance competitive with much larger models like ChatGPT and UniNER. Models can also be fine-tuned on your own data.

In a guardrails and PII context, GLiNER is widely used as a detection backend: you can define sensitive entity types (names, emails, locations) as labels and flag them in user or model text. It is, for instance, used inside Microsoft Presidio for PII detection.

// What it does

• Zero-shot NER: name any entity types at runtime, no retraining needed
• Lightweight, CPU-friendly models suited to consumer hardware
• Supports joint entity and relation extraction beyond plain NER
• Fine-tunable on your own labeled data
• Pretrained models distributed via Hugging Face (e.g. urchade/gliner_medium-v2.1)
• Optional Ray Serve deployment with dynamic batching and multi-replica scaling

// Getting started

Install the package, load a pretrained model, and pass your own entity labels to predict_entities.

pip install gliner

from gliner import GLiNER

model = GLiNER.from_pretrained("urchade/gliner_medium-v2.1")

text = "Cristiano Ronaldo was born on 5 February 1985 and plays for Al Nassr."

# Define the entity types you want to find
labels = ["Person", "Date", "Teams"]

entities = model.predict_entities(text, labels, threshold=0.5)

for entity in entities:
    print(entity["text"], "=>", entity["label"])

pip install gliner[serve]

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Detect and flag PII (names, emails, locations) in user input or model output as a guardrails backend
• Extract domain-specific entities from documents without labeling data for a fixed schema
• Run NER on CPU or consumer hardware where a large LLM is too costly
• Fine-tune a small NER model on your own entity types for an in-house pipeline

// How GLiNER compares

GLiNER alongside other open-source guardrails & security tools AI/TLDR tracks, ranked by GitHub stars.