bench: consolidate AIME export to suites only; set num_runs=8 default

Author: benjibc

eval_protocol/benchmarks/run.py

@@ -38,6 +38,8 @@ def _parse_args() -> argparse.Namespace:
         help="Limit rows: integer or 'all' for no limit (maps to EP_MAX_DATASET_ROWS)",
     )
     parser.add_argument("--num-runs", type=int, help="Override num_runs if provided")
+    # Allow overriding reasoning effort explicitly (low/medium/high). If omitted, suite default is used.
+    # Already mapped by --reasoning-effort above.
     return parser.parse_args()
 
 

eval_protocol/benchmarks/suites/aime25.py

@@ -74,7 +74,7 @@ def aime2025_dataset_adapter(rows: List[Dict[str, Any]]) -> List[EvaluationRow]:
     rollout_processor=default_single_turn_rollout_processor,
     aggregation_method="mean",
     threshold_of_success=None,
-    num_runs=1,
+    num_runs=8,
     max_dataset_rows=2,
     max_concurrent_rollouts=4,
     mode="pointwise",

examples/aime2025_chat_completion/tests/test_evaluation.py

@@ -6,7 +6,6 @@
     default_single_turn_rollout_processor,
 )
 from eval_protocol.pytest.evaluation_test import evaluation_test
-from eval_protocol.benchmarks.registry import export_benchmark
 
 from examples.aime2025_chat_completion.main import _extract_boxed_text, _normalize_to_int_or_none
 
@@ -56,7 +55,6 @@ def aime2025_dataset_adapter(rows: List[Dict[str, Any]]) -> List[EvaluationRow]:
     return converted
 
 
-@export_benchmark("aime25_low")
 @evaluation_test(
     model=["fireworks_ai/accounts/fireworks/models/gpt-oss-120b"],
     input_dataset=[