fix while break logic

code_eval.py

@@ -192,12 +192,15 @@ class CodeEval(evaluate.Metric):
 
             pbar = tqdm(total=len(results))
             prev_done_count = 0
-            while not all(r.done() for r in results.values()):
+            done = False
+            while not done:
+                done = True
                 cur_done_count = 0
                 for result in results.values():
                     for r in result:
                         if not r.done():
                             r.refresh(early_stop)
+                            done = False
                         else:
                             cur_done_count += 1
                 pbar.update(cur_done_count - prev_done_count)