spideystreet/DrBERT-MedicalNER-FR

NER DrBERT (FR Medical) - Fine-tuned Model

Available here: @Model

Model Description

This is a fine-tuned Named Entity Recognition (NER) model based on DrBERT-7GB, specifically trained for French medical text processing.
The model is designed to extract clinical entities from unstructured French medical notes.

Base Model

• Backbone: @DrBERT-7GB
• Architecture: RoBERTa-based transformer with 12 layers, 12 attention heads, 768 embedding dimension
• Training Data: NACHOS 7GB corpus (French medical text)

Fine-tuning Dataset

• Dataset: @TypicaAI/MedicalNER_Fr
• Task: Token Classification (NER)
• Language: French
• Domain: Medical/Clinical

Model Performance

Training Configuration

• Learning Rate: 3e-5
• Batch Size: 8
• Max Sequence Length: 384
• Training Epochs: 4
• Optimizer: AdamW
• Weight Decay: 0.01

Validation Metrics

• F1 Score: ~0.73 (span-level)
• Precision: ~0.75
• Recall: ~0.71
• Accuracy: ~0.95

Note: Metrics from early POC training. Performance may vary with different configurations.

Supported Entity Types

The model can identify the following medical entities:

Entity Type	Description	Example
PERSON	Patient names, medical personnel	"M. Dupont", "Dr. Martin"
SYMPTOM	Medical symptoms	"douleur", "fièvre", "nausée"
DISEASE	Medical conditions	"hypertension", "diabète"
MEDICATION	Drugs and treatments	"paracétamol", "insuline"
PROCEDURE	Medical procedures	"radiographie", "chirurgie"
ANATOMY	Body parts and structures	"cœur", "poumon", "abdomen"
LOCATION	Anatomical locations	"bras droit", "thorax"
ORGANIZATION	Medical institutions	"hôpital", "clinique"
CLINICAL_TERM	Clinical terminology	"diagnostic", "traitement"
GROUP	Patient groups	"enfants", "adultes"
PRODUCT	Medical products	"appareil", "équipement"

Model Limitations

Current Limitations

• Sequence Length: Limited to 384 tokens (longer notes are truncated)
• Class Imbalance: Some entity types may be underrepresented
• Domain Specificity: Optimized for French medical text
• Early POC: This is a proof-of-concept model, not production-ready

Known Issues

• May struggle with very long medical reports
• Performance may vary with different medical specialties
• Requires validation by medical professionals for clinical use

Training Details

Data Preprocessing

• Tokenization using DrBERT-7GB tokenizer
• Label alignment for subword tokens
• Train/validation/test split: 80/10/10
• Data augmentation: None (preserving medical accuracy)

Training Environment

• Hardware: Mac Air M3 16GB
• Framework: PyTorch with Hugging Face Transformers
• Training Time: ~2 hours for 4 epochs

Model Files

This directory contains:

• config.json - Model configuration
• model.safetensors - Model weights (SafeTensors format)
• tokenizer.json - Tokenizer configuration
• tokenizer_config.json - Tokenizer settings
• special_tokens_map.json - Special tokens mapping
• training_args.bin - Training arguments
• checkpoint-*/ - Training checkpoints

Citation

If you use this model, please cite:

@inproceedings{labrak2023drbert,
    title = {{DrBERT: A Robust Pre-trained Model in French for Biomedical and Clinical domains}},
    author = {Labrak, Yanis and Bazoge, Adrien and Dufour, Richard and Rouvier, Mickael and Morin, Emmanuel and Daille, Béatrice and Gourraud, Pierre-Antoine},
    booktitle = {Proceedings of the 61th Annual Meeting of the Association for Computational Linguistics (ACL'23), Long Paper},
    month = july,
    year = {2023},
    address = {Toronto, Canada},
    publisher = {Association for Computational Linguistics}
}

License

This model is released under the OpenRail License. See the LICENSE file in this directory for details.

The OpenRail License is designed for AI models and provides:

• Open use for research and commercial purposes
• Responsible AI guidelines and restrictions
• Attribution requirements
• Safety and ethical use provisions

Medical Use Disclaimer

⚠️ IMPORTANT: This model is for research and development purposes only. It is NOT approved for clinical use and should not be used for direct patient care without proper validation by qualified medical professionals.

Contact

• Project: @MediNotes
• Issues: GitHub Issues
• Author: @spidey

---

Last updated: September 2025 Model version: POC v0.1.1