Add validation to conversion script

convert_datasets_to_json.py

@@ -1,7 +1,13 @@
+import subprocess
+import urllib
+
+import typer
 from datasets import (get_dataset_config_names, get_dataset_split_names,
                       load_dataset)
 from huggingface_hub import list_datasets
 
+app = typer.Typer()
+
 
 def convert(dataset_id: str):
     dataset_name = dataset_id.split("/")[-1]
@@ -13,27 +19,57 @@ def convert(dataset_id: str):
         for split in splits:
             columns_to_keep = ["gem_id", "gem_parent_id", "target"]
             dataset = load_dataset(dataset_id, name=config, split=split)
-            # It seems like we store the references column as the target one
-            dataset = dataset.map(lambda x: {"target": x["references"]})
+            # For non-train splits, it seems we use the references column as the target
+            if "train" not in split:
+                dataset = dataset.map(lambda x: {"target": x["references"]})
+            else:
+                # Wrap references in list to match GEM schema
+                dataset = dataset.map(lambda x: {"target": [x["target"]]})
             # Delete unused columns
             # The test split doesn't have a parent ID
             if split == "test":
                 columns_to_keep.remove("gem_parent_id")
             # The `datasets` JSON serializer is buggy - use `pandas` for now
             df = dataset.to_pandas()
-            df[columns_to_keep].to_json(f"{dataset_name}_{config}_{split}.json", orient="records")
-            # TODO: validate against existing references on GitHub
-            # diff <(jq --sort-keys . mlsum_de_challenge_test_covid.json) <(jq --sort-keys . ~/git/GEM-metrics/data/references/mlsum_de_challenge_test_covid.json)
+            reference_name = f"{dataset_name}_{config}_{split}"
+            df[columns_to_keep].to_json(f"{reference_name}.json", orient="records")
+            validate(reference_name)
+
+
+def validate(reference_name: str):
+    # Download original references from GitHub repo
+    url = f"https://github.com/GEM-benchmark/GEM-metrics/releases/download/data/{reference_name}.json"
+    urllib.request.urlretrieve(
+        url,
+        f"github_references/{reference_name}.json",
+    )
+    # Run diff - requires `jq`
+    process = subprocess.run(
+        f"diff <(jq --sort-keys . {reference_name}.json) <(jq --sort-keys . ./github_references/{reference_name}.json)",
+        shell=True,
+        stdout=subprocess.PIPE,
+        executable="/bin/bash",
+    )
+    if process.stdout:
+        raise ValueError(f"❌ Validation failed for {reference_name}! New and original references do not agree 😭")
+    else:
+        typer.echo(f"✅ Validation successful for {reference_name}!")
 
 
+@app.command()
 def main():
     all_datasets = list_datasets()
+    # Filter for GEM datasets
     gem_datasets = [dataset for dataset in all_datasets if dataset.id.startswith("GEM/")]
-    # Test run with MLSUM
+    # Filter for blocklist
+    blocklist = ["test", "test-transform"]
+    gem_datasets = [dataset for dataset in gem_datasets if dataset.id not in blocklist]
+    # Focus on MLSUM for now - later remove this line to evaluate all datasets
     mlsum_datasets = [dataset for dataset in gem_datasets if dataset.id.startswith("GEM/mlsum")]
     for dataset in mlsum_datasets:
         convert(dataset.id)
+    typer.echo(f"🥳 All datasets converted!")
 
 
 if __name__ == "__main__":
-    main()
+    app()

requirements.txt

@@ -0,0 +1,3 @@
+datasets==1.16.1
+huggingface-hub==0.2.1
+typer==0.4.0