Update retrain_from_feedback.py

retrain_from_feedback.py

@@ -15,39 +15,42 @@ CSV_PATH = "feedback_log.csv"
 def train_evo():
     if not os.path.exists(CSV_PATH):
         print("⚠️ No feedback_log.csv file found.")
-        return
+        return "⚠️ No feedback data file found."
 
     df = pd.read_csv(CSV_PATH)
-    if df.empty:
-        print("⚠️ feedback_log.csv is empty.")
-        return
+
+    # ✅ Only use rows where vote is Evo or GPT
+    usable_df = df[df["vote"].isin(["Evo", "GPT"])].copy()
+
+    if usable_df.empty:
+        print("⚠️ No usable feedback data. Please vote on Evo or GPT.")
+        return "⚠️ No usable feedback data. Please vote on Evo or GPT."
 
     # Step 1: Evolve new architecture
     base_config = default_config()
     evolved_config = mutate_genome(base_config)
     print("🧬 New mutated config:", evolved_config)
 
-    # Step 2: Initialize model with evolved config
+    # Step 2: Initialize model
     model = EvoTransformerV22(
         num_layers=evolved_config["num_layers"],
         num_heads=evolved_config["num_heads"],
         ffn_dim=evolved_config["ffn_dim"],
         memory_enabled=evolved_config["memory_enabled"]
     )
-
     tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
     optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
     model.train()
 
-    # Step 3: Train on feedback
+    # Step 3: Train using feedback
     total_loss = 0.0
-    for _, row in df.iterrows():
+    for _, row in usable_df.iterrows():
         question = row["question"]
         opt1 = row["option1"]
         opt2 = row["option2"]
-        answer = row["answer"]
+        evo_answer = row["evo_answer"]
 
-        label = torch.tensor([1.0 if answer.strip() == opt2.strip() else 0.0])
+        label = torch.tensor([1.0 if evo_answer.strip() == opt2.strip() else 0.0])
 
         input_text = f"{question} [SEP] {opt2 if label.item() == 1 else opt1}"
         encoded = tokenizer(input_text, return_tensors="pt", padding="max_length", truncation=True, max_length=128)
@@ -59,12 +62,14 @@ def train_evo():
         optimizer.zero_grad()
         total_loss += loss.item()
 
-    # Step 4: Save new model weights
+    # Step 4: Save the retrained model
     torch.save(model.state_dict(), MODEL_PATH)
     print("✅ Evo model retrained and saved.")
 
-    # Step 5: Log genome and score (loss as proxy)
-    avg_loss = total_loss / len(df)
-    score = 1.0 - avg_loss  # Use (1 - loss) as crude fitness
-    log_genome(evolved_config, performance=round(score, 4))
-    print("🧬 Genome logged with score:", round(score, 4))
+    # Step 5: Log genome with fitness score (1 - avg_loss)
+    avg_loss = total_loss / len(usable_df)
+    fitness = round(1.0 - avg_loss, 4)
+    log_genome(evolved_config, score=fitness)
+    print("🧬 Genome logged with score:", fitness)
+
+    return f"✅ Evo retrained. Loss: {avg_loss:.4f}, Fitness: {fitness}"