davanstrien HF Staff commited on
Commit
e3e057f
·
1 Parent(s): a090310
Files changed (1) hide show
  1. semantic-dedupe.py +22 -24
semantic-dedupe.py CHANGED
@@ -14,6 +14,9 @@ Semantic deduplication for Hugging Face datasets using SemHash.
14
  This script removes duplicate or near-duplicate text samples from datasets based on
15
  semantic similarity, helping to clean training data and prevent train/test leakage.
16
 
 
 
 
17
  Example usage:
18
  # Basic deduplication
19
  uv run semantic-dedupe.py username/dataset text username/dataset-deduped
@@ -22,8 +25,8 @@ Example usage:
22
  uv run semantic-dedupe.py username/dataset text username/dataset-deduped \\
23
  --threshold 0.85 --max-samples 1000
24
 
25
- # Using HF Jobs with GPU
26
- hf jobs uv run --flavor a10 \\
27
  -e HF_TOKEN=$(python3 -c "from huggingface_hub import get_token; print(get_token())") \\
28
  https://huggingface.co/datasets/uv-scripts/deduplication/raw/main/semantic-dedupe.py \\
29
  username/dataset text username/dataset-deduped
@@ -202,44 +205,39 @@ def main():
202
  print(f"Available columns: {', '.join(dataset.column_names)}")
203
  sys.exit(1)
204
 
 
 
 
 
205
  # Initialize SemHash
206
- print("Initializing SemHash...")
207
- semhash = SemHash(
208
- batch_size=args.batch_size,
209
- show_progress=True,
210
- )
211
 
212
  # Perform deduplication
213
  print(f"Performing {args.method} deduplication on '{args.column}' column...")
214
 
215
  if args.method == "duplicates":
216
- result = semhash.deduplicate(
217
- dataset,
218
- text_column=args.column,
219
- threshold=args.threshold,
220
- )
221
  elif args.method == "outliers":
222
- result = semhash.filter_outliers(
223
- dataset,
224
- text_column=args.column,
225
- threshold=args.threshold,
226
- )
227
  elif args.method == "representatives":
228
- result = semhash.get_representatives(
229
- dataset,
230
- text_column=args.column,
231
- threshold=args.threshold,
232
- )
233
  else:
234
  raise ValueError(f"Unknown method: {args.method}")
235
 
236
- deduped_size = len(result)
 
 
 
 
 
237
 
238
  # Print statistics
239
  print(f"\nDeduplication complete!")
240
  print(f"Original size: {original_size:,}")
241
  print(f"Deduplicated size: {deduped_size:,}")
242
  print(f"Removed: {original_size - deduped_size:,} ({((original_size - deduped_size) / original_size) * 100:.1f}%)")
 
243
 
244
  # Create dataset card
245
  card = create_dataset_card(
@@ -253,7 +251,7 @@ def main():
253
 
254
  # Push to hub
255
  print(f"\nPushing to hub: {args.output_repo}")
256
- result.push_to_hub(
257
  args.output_repo,
258
  private=args.private,
259
  commit_message=f"Deduplicated using {args.method} method",
 
14
  This script removes duplicate or near-duplicate text samples from datasets based on
15
  semantic similarity, helping to clean training data and prevent train/test leakage.
16
 
17
+ SemHash is CPU-optimized and uses Model2Vec embeddings that are 500x faster on CPU
18
+ than traditional transformers. No GPU required!
19
+
20
  Example usage:
21
  # Basic deduplication
22
  uv run semantic-dedupe.py username/dataset text username/dataset-deduped
 
25
  uv run semantic-dedupe.py username/dataset text username/dataset-deduped \\
26
  --threshold 0.85 --max-samples 1000
27
 
28
+ # Using HF Jobs (CPU is sufficient)
29
+ hf jobs uv run --flavor cpu-4x-xlarge \\
30
  -e HF_TOKEN=$(python3 -c "from huggingface_hub import get_token; print(get_token())") \\
31
  https://huggingface.co/datasets/uv-scripts/deduplication/raw/main/semantic-dedupe.py \\
32
  username/dataset text username/dataset-deduped
 
205
  print(f"Available columns: {', '.join(dataset.column_names)}")
206
  sys.exit(1)
207
 
208
+ # Convert dataset to records (preserves all columns)
209
+ print("Converting dataset to records...")
210
+ records = [dict(row) for row in dataset]
211
+
212
  # Initialize SemHash
213
+ print("Initializing SemHash (CPU-optimized)...")
214
+ semhash = SemHash.from_records(records=records, columns=[args.column])
 
 
 
215
 
216
  # Perform deduplication
217
  print(f"Performing {args.method} deduplication on '{args.column}' column...")
218
 
219
  if args.method == "duplicates":
220
+ result = semhash.self_deduplicate(threshold=args.threshold)
 
 
 
 
221
  elif args.method == "outliers":
222
+ result = semhash.self_filter_outliers()
 
 
 
 
223
  elif args.method == "representatives":
224
+ result = semhash.self_find_representative()
 
 
 
 
225
  else:
226
  raise ValueError(f"Unknown method: {args.method}")
227
 
228
+ # Get deduplicated records (all columns preserved)
229
+ deduplicated_records = result.selected
230
+
231
+ # Convert back to HF Dataset
232
+ result_dataset = Dataset.from_list(deduplicated_records)
233
+ deduped_size = len(result_dataset)
234
 
235
  # Print statistics
236
  print(f"\nDeduplication complete!")
237
  print(f"Original size: {original_size:,}")
238
  print(f"Deduplicated size: {deduped_size:,}")
239
  print(f"Removed: {original_size - deduped_size:,} ({((original_size - deduped_size) / original_size) * 100:.1f}%)")
240
+ print("\nNote: SemHash processes ~20,000 sentences/second on CPU")
241
 
242
  # Create dataset card
243
  card = create_dataset_card(
 
251
 
252
  # Push to hub
253
  print(f"\nPushing to hub: {args.output_repo}")
254
+ result_dataset.push_to_hub(
255
  args.output_repo,
256
  private=args.private,
257
  commit_message=f"Deduplicated using {args.method} method",