Robust extraction of positive-class prob from r_probs (any shape)

app.py

@@ -132,9 +132,12 @@ def predict(sequence: str):
 
     pred = pred_texts[0]
     r = float(r_pred.cpu().tolist()[0] if torch.is_tensor(r_pred) else r_pred[0])
-    r_probs_list = r_probs.cpu().tolist() if torch.is_tensor(r_probs) else list(r_probs)
-    first = r_probs_list[0]
-    p_pos = float(first[1]) if isinstance(first, (list, tuple)) else float(r_probs_list[1])
+    if torch.is_tensor(r_probs):
+        flat = r_probs.flatten().cpu().tolist()
+    else:
+        flat = [float(x) for sub in r_probs for x in (sub if isinstance(sub, (list, tuple)) else [sub])]
+    print(f"[debug] r_probs raw flat = {flat}")  # remove after verifying
+    p_pos = float(flat[-1])
     return pred, format_reliability(r), f"{p_pos:.4f}"