nps-score-distilbert-multilingual

NPS Score Prediction — distilbert_multilingual

Fine-tuned distilbert-base-multilingual-cased pour prédire le score NPS sentimental (0–10) depuis les commentaires clients.
Projet : Système NPS Dior/Reetain — Nada El Maliki

Métriques (val set)

Usage

import torch
from transformers import AutoTokenizer
from modeling_nps_score import NPSScoreModel

model = NPSScoreModel.from_pretrained(
    "nada-05/nps-score-distilbert-multilingual", trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    "nada-05/nps-score-distilbert-multilingual", trust_remote_code=True
)
model.eval()

text = "J'ai attendu 40 minutes sans assistance. Vendeur peu aimable."
enc  = tokenizer(text, return_tensors="pt", max_length=256,
                 truncation=True, padding="max_length")
with torch.no_grad():
    out = model(**enc)
score_nps = round(out.logits.item() * 10, 1)  # → ex: 2.8
print(f"Score IA : {score_nps}/10")

Architecture

• Encoder : distilbert-base-multilingual-cased
• Tête : Linear(hidden→128) → GELU → Dropout → Linear(128→1) → Sigmoid
• Loss d'entraînement : MSE
• Dataset : 3018 réponses NPS FR/EN (2095 train / 449 val / 449 test)