---
language:
- fr
- en
tags:
- nps
- text-generation
- seq2seq
- customer-experience
- crm
license: mit
---

# NPS Description Generator (mT5-small)

Fine-tuned **google/mt5-small** pour générer des descriptions CRM personnalisées  
pour les clients détracteurs NPS. Projet : Système NPS Dior/Reetain — Nada El Maliki.

## Métriques
| Métrique | Valeur |
|----------|--------|
| ROUGE-2 (best) | 0.1529 |
| ROUGE-1 (best) | ~0.32 |
| ROUGE-L (best) | ~0.26 |

## Usage

```python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

tokenizer = AutoTokenizer.from_pretrained("nada-05/nps-description-generator-mt5")
model     = AutoModelForSeq2SeqLM.from_pretrained("nada-05/nps-description-generator-mt5")
model.eval()

# Construire le prompt
prompt = (
    "generate nps description: [lang:fr] "
    "score:3/10 segment:Detractor urgency:high "
    "comments: J'ai attendu 40 minutes sans assistance. "
    "improvements: advisor | store"
)

enc = tokenizer(prompt, return_tensors="pt", max_length=256, truncation=True)
with torch.no_grad():
    out = model.generate(**enc, max_new_tokens=200, num_beams=4,
                          no_repeat_ngram_size=3)
description = tokenizer.decode(out[0], skip_special_tokens=True)
print(description)
```

## Format du prompt
```
generate nps description: [lang:<fr|en>]
  score:<0-10>/10
  segment:<Detractor|Passive|Promoter>
  urgency:<high|medium|low|none>
  comments: <texte commentaires>
  improvements: <axes signalés>   (optionnel)
```

## Architecture
- Base : `google/mt5-small` (~300M params)
- Dataset : 1014 exemples détracteurs FR/EN (fine-tuning 10 epochs)
- Loss : Cross-Entropy seq2seq