T-AIA-911-camembert-ner

This model is a fine-tuned CamemBERT for Named Entity Recognition (NER) on French railway transport texts. It identifies departure/arrival cities, transport modes, temporal expressions, persons, and miscellaneous entities in French queries.

Model Details

Model Description

This model was trained on a corpus annotated by the PAR_2 group of the T-AIA-911 project (MSc Pro IA/DATA, Epitech). It is based on CamemBERT (camembert-base) and adapted for the NER task with a BIO scheme.

• Developed by: PAR_2 group, MSc Pro IA/DATA, Epitech
• Shared by: glopez-ai
• Model type: Token Classification (NER)
• Language(s): French
• License: MIT
• Finetuned from: camembert-base

Model Sources

• Repository: https://huggingface.co/glopez-ai/T-AIA-911-camembert-ner
• Dataset: https://huggingface.co/datasets/glopez-ai/T-AIA-911-travel-orders

Uses

Direct Use

Named entity extraction in French transport queries, e.g., for voice assistants, chatbots, or booking systems.

Downstream Use

Integration into dialogue pipelines, search engines, or mobility applications.

Out-of-Scope Use

• Texts unrelated to transport or outside the railway domain
• Languages other than French
• Use for high-impact decisions without human validation

Bias, Risks, and Limitations

• Potential biases related to the dataset (railway domain, informal style)
• Lower performance on formal texts or other domains
• May poorly detect rare or ambiguous entities

Recommendations

Always validate model outputs in critical contexts. Adapt or retrain for other domains.

How to Get Started with the Model

from transformers import pipeline

ner = pipeline(
    "ner",
    model="glopez-ai/T-AIA-911-camembert-ner",
    aggregation_strategy="simple"
)

sentence = "Je veux aller de Paris à Lyon demain matin en TGV"
print(ner(sentence))

Training Details

Training Data

• Data annotated by the PAR_2 group (T-AIA-911, MSc Pro IA/DATA, Epitech), from traveler queries (see dataset)
• BIO scheme, entities: DEP, ARR, TRANSPORT, TEMP, PER, MISC

Training Procedure

• Mixed precision: fp16
• Epochs: 3
• Batch size: 16
• Learning rate: 5e-5
• Warmup ratio: 0.1
• Weight decay: 0.01
• Scheduler: linear

Speeds, Sizes, Times

• ~3 minutes training on T4 GPU for 3 epochs
• Model size: ~440 MB

Evaluation

Testing Data, Factors & Metrics

• T-AIA-911 project test set (see dataset)
• Metrics: F1, precision, recall, exact match (via seqeval)

Results

• F1 (test): see training_results.json file in the repo/model

Summary

The model achieves good performance on the target domain but requires validation on other types of texts.

Model Examination

• Interpretability via HuggingFace pipeline, visualization of extracted entities

Environmental Impact

• Hardware Type: T4 GPU (Google Colab)
• Hours used: < 0.1h
• Cloud Provider: Google Colab
• Compute Region: Europe
• Carbon Emitted: Estimated < 0.1 kgCO2eq

Technical Specifications

Model Architecture and Objective

• CamemBERT (base) + token classification (NER)

Compute Infrastructure

• Hardware: T4 GPU
• Software: Python 3.10, PyTorch, Transformers, Datasets

Citation

BibTeX:

@misc{par2_2026_camembertner,
  title={T-AIA-911 CamemBERT NER},
  author={PAR_2 group, MSc Pro IA/DATA, Epitech},
  year={2026},
  howpublished={\url{https://huggingface.co/glopez-ai/T-AIA-911-camembert-ner}}
}

Model Card Authors

• PAR_2 group, MSc Pro IA/DATA, Epitech

Model Card Contact

• https://huggingface.co/glopez-ai

Let me know if you need further adjustments!