Upload app.py with huggingface_hub

app.py

@@ -121,11 +121,16 @@ def check_faithfulness(context, question, answer, model_choice):
         u_score = logits[0, unfaithful_ids[0]].item()
     latency = (time.time() - start_time) * 1000
 
+    # Calibrated threshold: require faithful logit to be 0.9 higher than unfaithful
+    # This improves balanced accuracy from 67% to 72% by catching more hallucinations
+    CALIBRATION_THRESHOLD = 0.9
+    margin = f_score - u_score
+
     scores = torch.tensor([f_score, u_score])
     probs = F.softmax(scores, dim=0)
     confidence = probs.max().item() * 100
 
-    if f_score > u_score:
+    if margin > CALIBRATION_THRESHOLD:
         verdict = "FAITHFUL"
         color = "#22c55e"
         explanation = "The answer appears to be supported by the provided context."