[Ref, Fix] use cached list of usable models, convert logits to OneHot for EC as well

src/dataloading.py

@@ -137,3 +137,7 @@ def load_run_data(model_name, dataset_name):
 def load_run_data_cached(model_name, dataset_name):
     return load_run_data(model_name, dataset_name)
 
+
+if __name__ == "__main__":
+    get_leaderboard_models_reload()
+    

src/similarity.py

@@ -38,6 +38,7 @@ def compute_pairwise_similarities(metric_name: str, probs: list[list[np.array]],
         # Convert probabilities to one-hot
         probs = [[one_hot(p) for p in model_probs] for model_probs in probs]
     elif metric_name == "Error Consistency":
+        probs = [[one_hot(p) for p in model_probs] for model_probs in probs]
         metric = EC()
     else:
         raise ValueError(f"Invalid metric: {metric_name}")