GLiNER Model Documentation

Introduction

GLiNER is a Named Entity Recognition (NER) model designed to identify various entity types using a bidirectional transformer encoder architecture. It offers a scalable alternative to traditional NER models and Large Language Models (LLMs), being both resource-efficient and flexible for recognizing an unlimited number of entity types.

Architecture

The GLiNER model employs a bi-encoder architecture:

Textual Encoder: ModernBERT-large
Entity Label Encoder: BGE-base-en

This setup allows for:

Recognition of unlimited entities simultaneously.
Faster inference when entity embeddings are preprocessed.
Improved generalization to unseen entities.

The architecture improves efficiency up to fourfold compared to DeBERTa-based models and supports context lengths of up to 8,192 tokens.

Training

The model has been trained on datasets including urchade/pile-mistral-v0.1, numind/NuNER, and knowledgator/GLINER-multi-task-synthetic-data. It demonstrates competitive benchmarking results across various NER datasets.

Guide: Running Locally

Install GLiNER and Transformers Library:

pip install gliner -U
pip install git+https://github.com/huggingface/transformers.git

Load and Use the Model:

from gliner import GLiNER

model = GLiNER.from_pretrained("knowledgator/modern-gliner-bi-large-v1.0")

text = "Your input text here"
labels = ["entity types"]
entities = model.predict_entities(text, labels, threshold=0.3)

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

Optional Enhancements:

Install flash attention for better performance:
```
pip install flash-attn triton
```

Pre-embed entities for large datasets:

entity_embeddings = model.encode_labels(labels, batch_size=8)
outputs = model.batch_predict_with_embeds(texts, entity_embeddings, labels)

Cloud GPUs: Consider using cloud services like AWS EC2, Google Cloud Platform, or Azure for GPU support to handle large-scale inference or training efficiently.

License

GLiNER is released under the Apache-2.0 License, allowing for broad use and modification with proper attribution.