diff --git a/private_run_specs.conf b/private_run_specs.conf
index 0ad6b947..f5ade0a3 100644
--- a/private_run_specs.conf
+++ b/private_run_specs.conf
@@ -1,3 +1,4 @@
+
 entries: [    
   ## Real
 
@@ -82,6 +83,8 @@ entries: [
   {description: "math:model=text_code,subject=counting_and_probability,level=5,use_chain_of_thought=True", priority: 3}
   {description: "math:model=text_code,subject=precalculus,level=5,use_chain_of_thought=True", priority: 3}
 
-    {description: "sam_sum:model=neurips/local,max_train_instances=3", priority: 1}
-    {description: "ethics_utilitarianism:model=neurips/local", priority: 1}
+  {description: "sam_sum:model=neurips/local,max_train_instances=3", priority: 1}
+  {description: "ethics_utilitarianism:model=neurips/local", priority: 1}
+  {description: "corr2cause:model=neurips/local", priority: 1}
+
 ]
\ No newline at end of file
diff --git a/src/helm/benchmark/run_specs.py b/src/helm/benchmark/run_specs.py
index b08d51bd..19cfb49d 100644
--- a/src/helm/benchmark/run_specs.py
+++ b/src/helm/benchmark/run_specs.py
@@ -1040,6 +1040,27 @@ def get_ethics_utilitarianism_spec(method: str = ADAPT_MULTIPLE_CHOICE_JOINT) ->
     )
 
 
+@run_spec_function("corr2cause")
+def get_corr2cause_spec(method: str = ADAPT_MULTIPLE_CHOICE_JOINT) -> RunSpec:
+    scenario_spec = ScenarioSpec(class_name="helm.benchmark.scenarios.corr2cause_scenario.Corr2CauseScenario", args={})
+
+    prompt = """
+Given a scenario with a premise and a hypothesis, determine if the hypothesis can be inferred from the premise.
+"""
+
+    adapter_spec = get_multiple_choice_adapter_spec(
+        method=method, max_tokens=1, instructions=prompt, input_noun="Scenario\n", output_noun="Answer"
+    )
+
+    return RunSpec(
+        name=f"corr2cause,method={method}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs(),
+        groups=["corr2cause"],
+    )
+
+
 @run_spec_function("twitter_aae")
 def get_twitter_aae_spec(demographic: str) -> RunSpec:
     scenario_spec = ScenarioSpec(
diff --git a/src/helm/benchmark/scenarios/corr2cause_scenario.py b/src/helm/benchmark/scenarios/corr2cause_scenario.py
new file mode 100644
index 00000000..b10d3d3b
--- /dev/null
+++ b/src/helm/benchmark/scenarios/corr2cause_scenario.py
@@ -0,0 +1,88 @@
+import csv
+import os
+import random
+from typing import List, Dict, Any
+from helm.common.general import ensure_file_downloaded, ensure_directory_exists
+from .scenario import Scenario, Instance, Reference, ALL_SPLITS, CORRECT_TAG, VALID_SPLIT, Input, Output
+
+# TODO: Should I just get rid of the train/test split?
+
+
+class Corr2CauseScenario(Scenario):
+    """Information on this class"""
+
+    name = "corr2cause"
+    description = "Can Large Language Models Infer Causation from Correlation?"
+    tags = ["classification"]
+    DATASET_FILE_NAME = "corr2cause.csv"
+    TRAIN_RATIO = 0.8  # 80% for training, 20% for validation
+    TRAIN_SPLIT = "train"
+    VALID_SPLIT = "valid"
+
+    def download_dataset(self, output_path: str):
+        """Downloads the Corr2Cause dataset if not already present."""
+        # Define the target path for the dataset
+        data_dir = os.path.join(output_path, "data")
+        dataset_path = os.path.join(data_dir, self.DATASET_FILE_NAME)
+
+        # Check if the dataset already exists
+        if os.path.exists(dataset_path):
+            print(f"The dataset '{self.DATASET_FILE_NAME}' already exists at '{dataset_path}'. Skipping download.")
+            return
+
+        # Download the raw data
+        url = "https://gist.githubusercontent.com/msaroufim/2835e9a27490bb183de86c54c0614169/raw/4160842cd2574716355a5fe9134387a20baed9f8/corr2cause.csv"
+        ensure_directory_exists(data_dir)
+        ensure_file_downloaded(source_url=url, target_path=dataset_path)
+
+    def load_dataset(self, output_path: str) -> List[Dict[str, Any]]:
+        file_path = os.path.join(output_path, "data", self.DATASET_FILE_NAME)
+
+        data = []
+        with open(file_path, encoding="utf-8") as f:
+            csv_reader = csv.reader(f)
+            next(csv_reader)
+            for label, question in csv_reader:
+                data_point = {"label": int(label), "input": question.strip()}
+                data.append(data_point)
+        random.seed(0)
+        random.shuffle(data)
+        return data
+    
+    def get_label(self, label: int) -> str:
+        return "No" if label == 0 else "Yes"
+    
+    def data_to_instance(self, data_point: Dict[str, Any], split: str, instance_id: str) -> Instance:
+        input_text = Input(text=data_point["input"])
+
+        # Create reference choices with "No" and "Yes"
+        choices = [
+            Reference(output=Output(text="No"), tags=[]),
+            Reference(output=Output(text="Yes"), tags=[])
+        ]
+
+        # Assign the CORRECT_TAG to the correct choice
+        correct_label = self.get_label(data_point["label"])
+        for choice in choices:
+            if choice.output.text == correct_label:
+                choice.tags.append(CORRECT_TAG)
+
+        return Instance(
+            input=input_text, references=choices, split=split
+        )
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        """Returns the instances for this scenario."""
+        self.download_dataset(output_path)
+        data = self.load_dataset(output_path)
+        # Split the data
+        split_k = int(len(data) * self.TRAIN_RATIO)
+        train_data = data[:split_k]
+        valid_data = data[split_k:]
+
+        train_instances = [self.data_to_instance(dt, self.TRAIN_SPLIT, f"id{i}") for i, dt in enumerate(train_data)]
+        valid_instances = [
+            self.data_to_instance(dt, self.VALID_SPLIT, f"id{i+len(train_data)}") for i, dt in enumerate(valid_data)
+        ]
+
+        return train_instances + valid_instances