diff --git a/environments/eval_environments/gsm8k_eval.py b/environments/eval_environments/gsm8k_eval.py index 8e53b7ef..9cacb5b4 100644 --- a/environments/eval_environments/gsm8k_eval.py +++ b/environments/eval_environments/gsm8k_eval.py @@ -353,7 +353,7 @@ class GSM8KEvalEnv(BaseEnv): # Create evaluation tasks async def eval_task(item): - return await self.rollout_and_score_eval(item, self.server_configs[0]) + return await self.rollout_and_score_eval(item, self.server.servers[0].config) tasks = [eval_task(item) for item in self.eval_items]