rename and improve comments

src/helm/benchmark/adaptation/adapters/adapter.py

@@ -40,9 +40,11 @@ def generate_requests(self, eval_instance: Instance) -> List[RequestState]:
         """
         pass
 
-    def sample_instances(self, instances: List[Instance]) -> List[Instance]:
+    def get_run_instances(self, instances: List[Instance]) -> List[Instance]:
         """
-        Leave the train instances alone. For the eval instances, keep at most `max_eval_instances`.
+        Get the instances necessary for this run:
+        Train instances (split=train): keep all (if any) for in-context learning
+        Eval instances (split=valid or test): keep at most `max_eval_instances` specified in `AdapterSpec` by sampling
         Return the resulting train and eval instances.
         """
         all_train_instances: List[Instance] = [instance for instance in instances if instance.split == TRAIN_SPLIT]

src/helm/benchmark/runner.py

@@ -106,7 +106,7 @@ def run_one(self, run_spec: RunSpec):
         run_path: str = os.path.join(self.runs_path, run_spec.name)
         ensure_directory_exists(run_path)
 
-        # Initialize the adapter
+        # Fetch and initialize the Adapter based on the `AdapterSpec`.
         adapter: Adapter = AdapterFactory.get_adapter(run_spec.adapter_spec, self.tokenizer_service)
 
         instances: List[Instance]
@@ -118,8 +118,8 @@ def run_one(self, run_spec: RunSpec):
             # Give each instance a unique ID
             instances = with_instance_ids(instances)
 
-            # Sample only as many as we need
-            instances = adapter.sample_instances(instances)
+            # Get the instances necessary for this run.
+            instances = adapter.get_run_instances(instances)
 
             # Data preprocessing
             instances = DataPreprocessor(run_spec.data_augmenter_spec).preprocess(