draft

semantic-dedupe.py

@@ -14,6 +14,9 @@ Semantic deduplication for Hugging Face datasets using SemHash.
 This script removes duplicate or near-duplicate text samples from datasets based on
 semantic similarity, helping to clean training data and prevent train/test leakage.
 
+SemHash is CPU-optimized and uses Model2Vec embeddings that are 500x faster on CPU
+than traditional transformers. No GPU required!
+
 Example usage:
     # Basic deduplication
     uv run semantic-dedupe.py username/dataset text username/dataset-deduped
@@ -22,8 +25,8 @@ Example usage:
     uv run semantic-dedupe.py username/dataset text username/dataset-deduped \\
         --threshold 0.85 --max-samples 1000
 
-    # Using HF Jobs with GPU
-    hf jobs uv run --flavor a10 \\
+    # Using HF Jobs (CPU is sufficient)
+    hf jobs uv run --flavor cpu-4x-xlarge \\
         -e HF_TOKEN=$(python3 -c "from huggingface_hub import get_token; print(get_token())") \\
         https://huggingface.co/datasets/uv-scripts/deduplication/raw/main/semantic-dedupe.py \\
         username/dataset text username/dataset-deduped
@@ -202,44 +205,39 @@ def main():
         print(f"Available columns: {', '.join(dataset.column_names)}")
         sys.exit(1)
     
+    # Convert dataset to records (preserves all columns)
+    print("Converting dataset to records...")
+    records = [dict(row) for row in dataset]
+    
     # Initialize SemHash
-    print("Initializing SemHash...")
-    semhash = SemHash(
-        batch_size=args.batch_size,
-        show_progress=True,
-    )
+    print("Initializing SemHash (CPU-optimized)...")
+    semhash = SemHash.from_records(records=records, columns=[args.column])
     
     # Perform deduplication
     print(f"Performing {args.method} deduplication on '{args.column}' column...")
     
     if args.method == "duplicates":
-        result = semhash.deduplicate(
-            dataset,
-            text_column=args.column,
-            threshold=args.threshold,
-        )
+        result = semhash.self_deduplicate(threshold=args.threshold)
     elif args.method == "outliers":
-        result = semhash.filter_outliers(
-            dataset,
-            text_column=args.column,
-            threshold=args.threshold,
-        )
+        result = semhash.self_filter_outliers()
     elif args.method == "representatives":
-        result = semhash.get_representatives(
-            dataset,
-            text_column=args.column,
-            threshold=args.threshold,
-        )
+        result = semhash.self_find_representative()
     else:
         raise ValueError(f"Unknown method: {args.method}")
     
-    deduped_size = len(result)
+    # Get deduplicated records (all columns preserved)
+    deduplicated_records = result.selected
+    
+    # Convert back to HF Dataset
+    result_dataset = Dataset.from_list(deduplicated_records)
+    deduped_size = len(result_dataset)
     
     # Print statistics
     print(f"\nDeduplication complete!")
     print(f"Original size: {original_size:,}")
     print(f"Deduplicated size: {deduped_size:,}")
     print(f"Removed: {original_size - deduped_size:,} ({((original_size - deduped_size) / original_size) * 100:.1f}%)")
+    print("\nNote: SemHash processes ~20,000 sentences/second on CPU")
     
     # Create dataset card
     card = create_dataset_card(
@@ -253,7 +251,7 @@ def main():
     
     # Push to hub
     print(f"\nPushing to hub: {args.output_repo}")
-    result.push_to_hub(
+    result_dataset.push_to_hub(
         args.output_repo,
         private=args.private,
         commit_message=f"Deduplicated using {args.method} method",