modif

scripts/inference.py

@@ -31,36 +31,45 @@ def zero_shot_inference(text):
 
 def few_shot_inference(text):
     """
-    Simule un few-shot learning en injectant des exemples dans le prompt (type prompt engineering).
+    Classification few-shot avec FLAN-T5 : génère uniquement le label (World, Sports, etc.).
     
     Args:
         text (str): Texte à classifier.
     
     Returns:
         tuple:
-            - str: Label prédit.
-            - dict: Scores pour chaque classe.
+            - str: Label prédit (nettoyé et validé).
+            - dict: Détails du texte généré brut.
     """
-    classifier = pipeline("zero-shot-classification", model="facebook/bart-large-mnli")
-    
-    #Exemples donnés au modèle pour le guider (prompt engineering)
+    model_name = "google/flan-t5-small"
+    classifier = pipeline("text2text-generation", model=model_name, max_new_tokens=10)
+
     examples = [
         ("The president met the UN delegation to discuss global peace.", "World"),
         ("The football team won their match last night.", "Sports"),
         ("The company reported a big profit this quarter.", "Business"),
         ("New research in AI shows promising results.", "Sci/Tech")
     ]
-    
-    #Construction du prompt avec des exemples
-    prompt = ""
-    for example_text, example_label in examples:
-        prompt += f"Text: {example_text}\nLabel: {example_label}\n\n"
-    prompt += f"Text: {text}\nLabel:"
 
-    candidate_labels = list(id2label.values())
-    result = classifier(prompt, candidate_labels)
-    prediction = result["labels"][0]
-    scores = {label: float(f"{score:.4f}") for label, score in zip(result["labels"], result["scores"])}
+    # Prompt few-shot
+    prompt = "Classify the following text into one of the following categories: World, Sports, Business, Sci/Tech.\n\n"
+    for ex_text, ex_label in examples:
+        prompt += f"Text: {ex_text}\nCategory: {ex_label}\n\n"
+    prompt += f"Text: {text}\nCategory:"
+
+    # Génération
+    output = classifier(prompt)[0]["generated_text"].strip()
+
+    # Nettoyage du label
+    output_clean = output.split()[0].rstrip(".").capitalize()  # ex : "sci/tech." → "Sci/tech"
+
+    # Mapping pour être sûr que ça correspond à une catégorie connue
+    candidate_labels = ["World", "Sports", "Business", "Sci/Tech"]
+    prediction = next((label for label in candidate_labels if label.lower() in output_clean.lower()), "Unknown")
+
+    # Fausse distribution (1.0 pour la classe prédite, 0.0 pour les autres)
+    scores = {label: 1.0 if label == prediction else 0.0 for label in candidate_labels}
+
     return prediction, scores
 
 

scripts/metric.py

@@ -1,5 +1,4 @@
 import time
-import numpy as np
 import pandas as pd
 from tqdm import tqdm  # ✅ Ajout ici
 from datasets import load_dataset
@@ -20,7 +19,7 @@ models_to_evaluate = {
 label_map = {0: "World", 1: "Sports", 2: "Business", 3: "Sci/Tech"}
 
 # Charger un sous-ensemble du jeu de test AG News
-dataset = load_dataset("ag_news", split="test[:10%]")  
+dataset = load_dataset("ag_news", split="test[:3]")  
 
 def evaluate_model(name, inference_func):
     print(f"\n🔍 Évaluation du modèle : {name}")
@@ -56,7 +55,7 @@ def evaluate_model(name, inference_func):
     f1 = f1_score(true_labels, pred_labels, average='weighted')
     prec = precision_score(true_labels, pred_labels, average='weighted')
     rec = recall_score(true_labels, pred_labels, average='weighted')
-    loss = log_loss(true_labels, all_probs)
+    loss = log_loss(true_labels, all_probs, labels=[0, 1, 2, 3])
 
     print(f"✅ Résultats {name} :")
     print(f"- Accuracy : {acc:.4f}")