The meatadata_*.json files under discoverybench/real/test do not seem to contain labeled hypothesis. How should we evaluate this portion of the dataset to get HMS scores?