Fix deduplication logic

project/nanoeval/nanoeval/evaluation.py

@@ -43,26 +43,16 @@ async def _make_pbar(spec: EvalSpec, n_tasks: int, recorder: RecorderProtocol) -
 
 
 def _create_clean_results(results: list[tuple[Task, Any]]) -> list[tuple[Task, Any]]:
-    """
-    Clean results are defined as the LAST retry_idx for each task. These results are what get
-    used to compute metrics in the final summary.
-    """
+    """Return only the latest attempt for each task."""
 
-    # Pick the latest retry idx for each task
-    clean_results: list[tuple[Task, Any]] = []
+    latest: dict[tuple[str, int], tuple[Task, Any]] = {}
     for task, result in results:
-        found = False
-        for i, (clean_task, _) in enumerate(clean_results):
-            if (clean_task.question_id, clean_task.attempt_id) == (
-                task.question_id,
-                task.attempt_id,
-            ):
-                found = True
-                if task.retry_idx > clean_task.retry_idx:
-                    clean_results[i] = (task, result)
-                break
-        if not found:
-            clean_results.append((task, result))
+        key = (task.question_id, task.attempt_id)
+        current = latest.get(key)
+        if current is None or task.retry_idx > current[0].retry_idx:
+            latest[key] = (task, result)
+
+    clean_results = list(latest.values())
 
     # Sanity check the deduplication function.
     seen_tasks: set[tuple[str, int]] = set()