try 1 concurrent rollout

xzrderek · xzrderek · commit c091e5be66be · 2025-10-14T22:51:29.000-07:00
diff --git a/eval_protocol/benchmarks/test_aime25.py b/eval_protocol/benchmarks/test_aime25.py
@@ -99,7 +99,7 @@ def aime2025_dataset_adapter(rows: List[Dict[str, Any]]) -> List[EvaluationRow]:
     passed_threshold=0.8,
     num_runs=1,
     max_dataset_rows=30,
-    max_concurrent_rollouts=2,
+    max_concurrent_rollouts=1,
     mode="pointwise",
 )
 def test_aime25_pointwise(row: EvaluationRow) -> EvaluationRow:

Original file line number	Diff line number	Diff line change
`@@ -99,7 +99,7 @@ def aime2025_dataset_adapter(rows: List[Dict[str, Any]]) -> List[EvaluationRow]:`
`99`	`99`	`passed_threshold=0.8,`
`100`	`100`	`num_runs=1,`
`101`	`101`	`max_dataset_rows=30,`
`102`		`- max_concurrent_rollouts=2,`
	`102`	`+ max_concurrent_rollouts=1,`
`103`	`103`	`mode="pointwise",`
`104`	`104`	`)`
`105`	`105`	`def test_aime25_pointwise(row: EvaluationRow) -> EvaluationRow:`