add gsm8k

tlem.py

@@ -77,6 +77,14 @@ class ReasoningMetric(evaluate.Metric):
         return results
 
 
+gsm8k = Task(
+    dataset_name=("gsm8k", "main"),
+    metric_name=("sustech/tlem", "gsm8k"),
+    input_column="question",
+    label_column="answer",
+)
+
+
 class Suite(EvaluationSuite):
     def run(
         self,
@@ -110,6 +118,8 @@ class Suite(EvaluationSuite):
                 suite = MMLU.suite(chat=chat)
             case _ if name.startswith("cmmlu"):
                 suite = CMMLU.suite(chat=chat)
+            case "gsm8k":
+                suite = [gsm8k]
         match name:
             case _ if "test" in name:
                 suite = suite["Test"]
@@ -120,20 +130,7 @@ class Suite(EvaluationSuite):
         super().__init__(name)
         self.cached_result = {}
 
-        match self.name:
-            case "cmmlu":
-                pass
-
         self.suite = [
-            Task(
-                dataset_name=("gsm8k", "main"),
-                metric_name=("sustech/tlem", "gsm8k"),
-                input_column="question",
-                label_column="answer",
-            )
             # TASK_REGISTRY["gsm8k"],
             # TASK_REGISTRY["competition_math"],
         ]
-
-
-# %%