lower max_new_tokens, generation takes too long

mlfoundations · Jan 19, 2025 · 850402c · 850402c
1 parent 23cad3a
commit 850402c
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 2 deletions.
diff --git a/eval/chat_benchmarks/AIME24/eval_instruct.py b/eval/chat_benchmarks/AIME24/eval_instruct.py
@@ -38,7 +38,7 @@ def __init__(
         super().__init__(logger)
         self.data_file = data_file
         self.debug = debug
-        self.max_new_tokens = 16384  # Setting this high to avoid truncation for reasoning models
+        self.max_new_tokens = 2048  # set higher to avoid truncation for reasoning models
 
     def generate_responses(self, model: LM) -> Dict[str, Any]:
         """

diff --git a/eval/chat_benchmarks/AMC23/eval_instruct.py b/eval/chat_benchmarks/AMC23/eval_instruct.py
@@ -39,7 +39,7 @@ def __init__(
         super().__init__(logger)
         self.data_file = data_file
         self.debug = debug
-        self.max_new_tokens = 16384  # Setting this high to avoid truncation for reasoning models
+        self.max_new_tokens = 2048  # set higher to avoid truncation for reasoning models
 
     def generate_responses(self, model: LM) -> Dict[str, Any]:
         """