aablaess/SAV-xlm-Roberta

🤖 XLM-RoBERTa — Classification d'Intentions Multilingue (SAV)

📋 Description

Ce modèle est un fine-tuning de XLM-RoBERTa base sur un dataset multilingue (Anglais, Français, Arabe) d'intentions clients dans le contexte du support client (SAV). Il est capable de reconnaître 27 intentions différentes exprimées dans les trois langues, et constitue le cœur du système de classification d'un assistant IA vocal.

---

🏆 Performance

Métrique	Epoch 1	Epoch 2	Epoch 3
Accuracy	99.73%	99.83%	99.88%
F1 Score (macro)	0.9973	0.9983	0.9988
Précision (macro)	0.9958	0.9975	0.9979
Rappel (macro)	0.9958	0.9974	0.9979
Validation Loss	0.0290	0.0142	0.0119

Le modèle atteint 99.88% d'accuracy sur le dataset de validation après 3 epochs, sans signe de surapprentissage (validation loss décroissante à chaque epoch).

---

🗂️ Dataset

Paramètre	Valeur
Source	Bitext Customer Support Dataset (27K)
Taille totale	~80 616 exemples (après extension multilingue)
Langues	Anglais 🇬🇧 · Français 🇫🇷 · Arabe 🇸🇦
Classes (intents)	27
Split entraînement	70% — 56 431 exemples
Split validation	15% — 12 092 exemples
Split test	15% — 12 093 exemples

🌍 Extension Multilingue

Le dataset source original est le Bitext Customer Support Dataset en anglais (~26 872 exemples). Il a été étendu vers le français et l'arabe via un pipeline de traduction automatique développé sur mesure :

• Traduction : bibliothèque deep_translator (GoogleTranslator) avec gestion du rate limiting (time.sleep(0.05) entre chaque ligne) et mécanisme de fallback (try/except) pour garantir l'intégrité des données en cas d'échec de traduction.
• Fichiers générés :
  • cleaned_bitext_dataset.csv — dataset anglais nettoyé (base)
  • dataset_fr_complete.csv — version française (~26 872 exemples)
  • dataset_ar_complete.csv — version arabe (~26 872 exemples)
  • dataset_multilingual_complete.csv — corpus unifié final (~80 616 exemples)
• Contrôle qualité : vérification du volume (tolérance de moins de 100 lignes d'écart entre source et traductions) avant fusion.
• Colonne language ajoutée automatiquement (en, fr, ar) pour la traçabilité.

Le corpus final de ~80 616 exemples (3 langues × ~26 872) a été utilisé pour entraîner le modèle, lui conférant une robustesse multilingue native.

---

🎯 Intentions Supportées (27 classes)

Catégorie	Intentions
ACCOUNT	create_account, delete_account, edit_account, switch_account, recover_password, registration_problems
ORDER	place_order, cancel_order, change_order, track_order
SHIPPING	delivery_options, delivery_period, change_shipping_address, set_up_shipping_address
REFUND	get_refund, track_refund, check_refund_policy, check_cancellation_fee
INVOICE	get_invoice, check_invoice
PAYMENT	check_payment_methods, payment_issue
CONTACT	contact_customer_service, contact_human_agent
AUTRE	complaint, review, newsletter_subscription

---

⚙️ Paramètres d'Entraînement

Paramètre	Valeur
Modèle de base	xlm-roberta-base
Learning rate	2e-5
Batch size	8 par GPU
Epochs	3
Max length	128 tokens
Weight decay	0.01
Précision	fp16 (demi-précision)
Environnement	Kaggle — GPU T4 x2
Durée d'entraînement	~88 minutes

---

🚀 Utilisation

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import pickle
from huggingface_hub import hf_hub_download

# Charger le modèle et le tokenizer
tokenizer = AutoTokenizer.from_pretrained("aablaess/SAV-xlm-Roberta")
model = AutoModelForSequenceClassification.from_pretrained("aablaess/SAV-xlm-Roberta")

# Charger le label encoder
label_encoder_path = hf_hub_download(
    repo_id="aablaess/SAV-xlm-Roberta",
    filename="label_encoder.pkl"
)
with open(label_encoder_path, "rb") as f:
    label_encoder = pickle.load(f)

# Prédiction
def predict_intent(text):
    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        padding=True,
        max_length=128
    )
    with torch.no_grad():
        outputs = model(**inputs)

    predicted_class = torch.argmax(outputs.logits, dim=1).item()
    intent = label_encoder.inverse_transform([predicted_class])[0]
    confidence = torch.softmax(outputs.logits, dim=1).max().item()

    return intent, confidence

# Exemples multilingues
texts = [
    "I want to cancel my order",        # Anglais
    "Je veux annuler ma commande",       # Français
    "أريد إلغاء طلبي"                   # Arabe
]

for text in texts:
    intent, confidence = predict_intent(text)
    print(f"Texte     : {text}")
    print(f"Intention : {intent} (confiance : {confidence:.2%})\n")

---

📁 Fichiers du Modèle

Fichier	Description
model.safetensors	Poids du modèle fine-tuné (1.11 GB)
config.json	Architecture du modèle (27 labels)
tokenizer.json	Vocabulaire et règles de tokenisation
tokenizer_config.json	Configuration du tokenizer
label_encoder.pkl	Correspondance index ↔ nom d'intention

---

📄 Licence

Ce modèle est distribué sous licence MIT.

---

---

🤖 XLM-RoBERTa — Multilingual Intent Classification (Customer Support)

📋 Description

This model is a fine-tuned version of XLM-RoBERTa base on a multilingual dataset (English, French, Arabic) of customer support intents. It can recognize 27 different intents expressed across three languages, and serves as the core classification engine of an AI voice assistant for customer service (SAV).

---

🏆 Performance

Metric	Epoch 1	Epoch 2	Epoch 3
Accuracy	99.73%	99.83%	99.88%
F1 Score (macro)	0.9973	0.9983	0.9988
Precision (macro)	0.9958	0.9975	0.9979
Recall (macro)	0.9958	0.9974	0.9979
Validation Loss	0.0290	0.0142	0.0119

The model reaches 99.88% accuracy on the validation dataset after 3 epochs, with no sign of overfitting (validation loss decreasing at each epoch).

---

🗂️ Dataset

Parameter	Value
Source	Bitext Customer Support Dataset (27K)
Total size	~80,616 examples (after multilingual extension)
Languages	English 🇬🇧 · French 🇫🇷 · Arabic 🇸🇦
Classes (intents)	27
Training split	70% — 56,431 examples
Validation split	15% — 12,092 examples
Test split	15% — 12,093 examples

🌍 Multilingual Extension

The original source dataset is the Bitext Customer Support Dataset in English (~26,872 examples). It was extended to French and Arabic using a custom-built automated translation pipeline:

• Translation: deep_translator library (GoogleTranslator) with rate limiting management (time.sleep(0.05) between each row) and a fallback mechanism (try/except) to ensure data integrity in case of translation failure.
• Generated files:
  • cleaned_bitext_dataset.csv — cleaned English dataset (base)
  • dataset_fr_complete.csv — French version (~26,872 examples)
  • dataset_ar_complete.csv — Arabic version (~26,872 examples)
  • dataset_multilingual_complete.csv — final unified corpus (~80,616 examples)
• Quality control: volume verification (tolerance of less than 100 rows difference between source and translations) before merging.
• language column automatically added (en, fr, ar) for traceability.

The final corpus of ~80,616 examples (3 languages × ~26,872) was used to train the model, giving it native multilingual robustness.

---

🎯 Supported Intents (27 classes)

Category	Intents
ACCOUNT	create_account, delete_account, edit_account, switch_account, recover_password, registration_problems
ORDER	place_order, cancel_order, change_order, track_order
SHIPPING	delivery_options, delivery_period, change_shipping_address, set_up_shipping_address
REFUND	get_refund, track_refund, check_refund_policy, check_cancellation_fee
INVOICE	get_invoice, check_invoice
PAYMENT	check_payment_methods, payment_issue
CONTACT	contact_customer_service, contact_human_agent
OTHER	complaint, review, newsletter_subscription

---

⚙️ Training Parameters

Parameter	Value
Base model	xlm-roberta-base
Learning rate	2e-5
Batch size	8 per GPU
Epochs	3
Max length	128 tokens
Weight decay	0.01
Precision	fp16 (half-precision)
Environment	Kaggle — GPU T4 x2
Training duration	~88 minutes

---

🚀 Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import pickle
from huggingface_hub import hf_hub_download

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("aablaess/SAV-xlm-Roberta")
model = AutoModelForSequenceClassification.from_pretrained("aablaess/SAV-xlm-Roberta")

# Load label encoder
label_encoder_path = hf_hub_download(
    repo_id="aablaess/SAV-xlm-Roberta",
    filename="label_encoder.pkl"
)
with open(label_encoder_path, "rb") as f:
    label_encoder = pickle.load(f)

# Predict intent
def predict_intent(text):
    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        padding=True,
        max_length=128
    )
    with torch.no_grad():
        outputs = model(**inputs)

    predicted_class = torch.argmax(outputs.logits, dim=1).item()
    intent = label_encoder.inverse_transform([predicted_class])[0]
    confidence = torch.softmax(outputs.logits, dim=1).max().item()

    return intent, confidence

# Multilingual examples
texts = [
    "I want to cancel my order",        # English
    "Je veux annuler ma commande",       # French
    "أريد إلغاء طلبي"                   # Arabic
]

for text in texts:
    intent, confidence = predict_intent(text)
    print(f"Text      : {text}")
    print(f"Intent    : {intent} (confidence: {confidence:.2%})\n")

---

📁 Model Files

File	Description
model.safetensors	Fine-tuned model weights (1.11 GB)
config.json	Model architecture (27 labels)
tokenizer.json	Vocabulary and tokenization rules
tokenizer_config.json	Tokenizer configuration
label_encoder.pkl	Index ↔ intent name mapping

---

📄 License

This model is distributed under the MIT license.