French environmental complaint classifier (Map Action, v2)

Classifies a short French complaint ("Le caniveau est bouch茅 devant l'茅cole") into the incident taxonomy used by the Map Action platform: 8 families and 32 incident types. Fine-tuned from intfloat/multilingual-e5-base (MIT).

Two models are in this repository: modele-famille/ (8 classes) and modele-type/ (32 classes). Prefix the input with query: .

Results

512 held-out texts. The test set only contains complaint formulations that never appear in training.

Version Families: macro F1 Families: accuracy Types: macro F1 Types: accuracy
TF-IDF + logistic regression 51.8 % 54.3 % 35.1 % 36.9 %
Frozen E5 embeddings + logistic regression 62.6 % 65.4 % 53.9 % 56.3 %
Fine-tuned E5, dataset v1 (2,048 training texts) 67.8 % 69.0 % 52.7 % 55.3 %
Fine-tuned E5, dataset v2 (3,328 training texts) 77.2 % 77.7 % 65.7 % 68.4 %

Training

8 epochs, batch 32, learning rate 4e-5, linear warm-up and decay. Data: Map-Action/plaintes-environnement-fr.

Limitations, please read

  • The data is synthetic. Complaints were written for each incident type and expanded with openings, places and consequences.
  • Some families overlap by design ("bad smells" appears under air quality and under nuisances), which caps the score.
  • In the full pipeline the input is a machine translation from Bambara, which adds its own errors.

Taxonomy

Eau & Assainissement 路 D茅chets & Insalubrit茅 路 Qualit茅 de l'Air 路 Risques Naturels & Climatiques 路 D茅gradation des Terres 路 Activit茅s mini猫res & extractives 路 Biodiversit茅 & 脡cosyst猫mes 路 Nuisances & Cadre de vie. The 32 types are listed in taxonomie.json.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 馃檵 Ask for provider support

Model tree for Map-Action/plaintes-fr-classifieur-v2

Finetuned
(194)
this model