Add fix_dataset.py utility script

fix_dataset.py

@@ -0,0 +1,96 @@
+#!/usr/bin/env -S uv run
+# /// script
+# requires-python = ">=3.10"
+# dependencies = [
+#     "datasets",
+#     "huggingface_hub",
+# ]
+# ///
+"""
+Fix dataset by deleting a split with mismatched schema and re-pushing.
+
+Usage:
+    uv run fix_dataset.py --repo evalstate/mcp-clients --split deduplicated
+"""
+
+import argparse
+import sys
+
+from datasets import load_dataset
+from huggingface_hub import HfApi
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Delete and recreate a dataset split")
+    parser.add_argument("--repo", required=True, help="Repository ID")
+    parser.add_argument("--split", required=True, help="Split name to fix")
+    parser.add_argument("--token", default=None, help="HF token")
+    args = parser.parse_args()
+
+    api = HfApi(token=args.token)
+
+    # Try to get the current repo info
+    try:
+        repo_info = api.repo_info(args.repo, repo_type="dataset")
+        print(f"Found repo: {args.repo}", file=sys.stderr)
+    except Exception as e:
+        print(f"Repo not found or error: {e}", file=sys.stderr)
+        sys.exit(1)
+
+    # The datasets library doesn't have a direct delete split API
+    # But we can use the deletion parameter in push_to_hub
+    # First, let's try to get the existing dataset and see what splits exist
+    
+    try:
+        ds = load_dataset(args.repo)
+        print(f"Existing splits: {list(ds.keys())}", file=sys.stderr)
+    except Exception as e:
+        print(f"Could not load dataset: {e}", file=sys.stderr)
+        sys.exit(1)
+
+    if args.split not in ds:
+        print(f"Split '{args.split}' not found in dataset", file=sys.stderr)
+        sys.exit(1)
+
+    # Delete the split by not including it in the push
+    # But first we need to save without that split and re-push
+    print(f"Deleting split '{args.split}'...", file=sys.stderr)
+    
+    # Use the deletions parameter to remove the split
+    try:
+        # Get all other splits
+        other_splits = {k: v for k, v in ds.items() if k != args.split}
+        
+        if other_splits:
+            print(f"Pushing without '{args.split}' split...", file=sys.stderr)
+            # Push only the other splits
+            other_splits['_all'] = other_splits.pop(list(other_splits.keys())[0])
+            
+        # Actually, the easiest way is to use hf CLI to delete files
+        # Let's find the parquet file for this split
+        print(f"Checking for split files to delete...", file=sys.stderr)
+        
+        # List all files in the repo
+        repo_files = api.list_repo_files(args.repo, repo_type="dataset")
+        split_files = [f for f in repo_files if f.startswith(f"data/{args.split}-") or f"split_{args.split}" in f]
+        
+        print(f"Files to delete for split '{args.split}': {split_files}", file=sys.stderr)
+        
+        if split_files:
+            print(f"Deleting {len(split_files)} files...", file=sys.stderr)
+            for f in split_files:
+                print(f"  Deleting: {f}", file=sys.stderr)
+                api.delete_file(f, args.repo, repo_type="dataset")
+            print("Files deleted. You'll need to re-push the split with correct schema.", file=sys.stderr)
+        else:
+            print("No split files found to delete", file=sys.stderr)
+            
+    except Exception as e:
+        print(f"Error: {e}", file=sys.stderr)
+        sys.exit(1)
+
+    print("Done! Re-run the dedup script to create the split with correct schema.", file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()