ALBERT-base-v2-French-NER (ONNX Optimized)

This model is a fine-tuned version of ALBERT-base-v2 for Named Entity Recognition (NER) on French text, using the wikiner_fr dataset.

It is designed to be ultra-lightweight (approximately 45 MB) and fast, making it ideal for edge applications or environments with limited CPU resources.

Performance

The model was trained for 2 epochs and achieves the following scores on the validation set:

• F1-Score: 0.8303
• Precision: 0.8306
• Recall: 0.8299
• Accuracy: 97.54%

Supported Entities

The model detects 4 types of entities:

• PER: Persons
• LOC: Locations
• ORG: Organizations
• MISC: Miscellaneous entities

Usage and Source Code

The complete code for training, evaluation, and inference (including optimized ONNX export) is available on GitHub:

Juste-Leo2/ALBERT-base-v2-french-ner

Quick Inference Example

from transformers import pipeline

# Direct loading via Transformers
ner_pipeline = pipeline(
    "token-classification", 
    model="JusteLeo/ALBERT-base-v2-french-ner",
    aggregation_strategy="simple"
)

text = "Bonjour, mon prénom est Thomas et j'habite à Paris."
results = ner_pipeline(text)

for entity in results:
    print(f"{entity['entity_group']} : {entity['word']} ({entity['score']:.2%})")

ONNX Optimization

The model is provided with an ONNX Runtime compatible configuration for increased execution speed on modern processors. You can find optimized conversion and inference scripts in the GitHub repository mentioned above.

---

License: Apache 2.0 Author: JusteLeo