rarmingaud/glidre_multi_v2

GLiDRE: Generalist and Lightweight Model for Document Relation Extraction - Multilingual checkpoint

GLiDRE v2

Better data, better training code

Key Features

• Zero-Shot Extraction: Capable of classifying unseen relations directly from text.
• Versatile Input Handling: Compatible with both tokenized text and full documents.
• Customizable Architecture: Supports multiple loss functions and allows easy modification of model components.

Installation

Install GLiDRE

pip install .

Quick Start

Here's a simple Python example to get you started:

from glidre import GLiDRE
model = GLiDRE.from_pretrained("cea-list-ia/glidre_multi_v2")
text = "The Loud Tour was the fourth overall and third world concert tour by Barbadian recording artist Rihanna."
# Define relation labels
labels = ["COUNTRY_OF_CITIZENSHIP", "PUBLICATION_DATE", "PART_OF"] # Labels are uppercase because the model performs better with capitalized relation names
# Define entity mentions (format: [{"id" : id, "type" : type, "mentions" : [{"value" : text, "start" : start_idx, "end" : end_idx}]}])
mentions = [{
                "id": 0,
                "mentions": [
                    {
                        "value": "Barbadian",
                        "start": 69,
                        "end": 78
                    }
                ],
                "type": "LOC"
            },
            {
                "id": 1,
                "mentions": [
                    {
                        "value": "Rihanna",
                        "start": 96,
                        "end": 103
                    }
                ],
                "type": "PER"}]
# Predict relations using GLiDRE
relations = model.predict_entities(text = text, labels = labels, mentions = mentions, threshold=0.3, multi_label = False)
print("Predicted Relations:")
for relation in relations:
    print(relation["entity_1"])
    print("Label :",  relation["relation_type"])
    print(relation["entity_2"])
    print("---")

Training

GLiDRE supports training on various datasets such as DocRED and Re-DocRED.

# For Re-DocRED:
python3 train.py --config configs/config_finetuning.yaml

Citation

@misc{armingaud2025glidregeneralistlightweightmodel,
      title={GLiDRE: Generalist Lightweight model for Document-level Relation Extraction}, 
      author={Robin Armingaud and Romaric Besançon},
      year={2025},
      eprint={2508.00757},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.00757}, 
}