fix: Use sink_parquet for true streaming (minimal memory)

long-context-pdfs.py

@@ -29,31 +29,11 @@ Examples:
 """
 
 import argparse
-import polars as pl
-from datasets import Features, IterableDataset, Value
-
-FEATURES = Features(
-    {
-        "id": Value("string"),
-        "url": Value("string"),
-        "text": Value("string"),
-        "language": Value("string"),
-        "token_count": Value("int64"),
-        "dump": Value("string"),
-        "page_average_lid_score": Value("float64"),
-    }
-)
-
+import tempfile
+from pathlib import Path
 
-def polars_to_generator(lf: pl.LazyFrame, chunk_size: int = 100):
-    """Stream LazyFrame as row generator with controlled memory.
-
-    Args:
-        lf: LazyFrame to stream
-        chunk_size: Rows per batch (lower = less memory). Default 100 for large text docs.
-    """
-    for batch_df in lf.collect_batches(chunk_size=chunk_size):
-        yield from batch_df.iter_rows(named=True)
+import polars as pl
+from datasets import Dataset
 
 
 def main():
@@ -81,18 +61,9 @@ def main():
     parser.add_argument("--limit", type=int, help="Limit rows")
     parser.add_argument("--output", type=str, help="Output dataset repo")
     parser.add_argument("--private", action="store_true")
-    parser.add_argument(
-        "--chunk-size",
-        type=int,
-        default=100,
-        help="Rows per batch (lower = less memory, default: 100)",
-    )
 
     args = parser.parse_args()
 
-    # Reduce streaming chunk size for large text documents (default can cause OOM)
-    pl.Config.set_streaming_chunk_size(1000)
-
     source = f"hf://datasets/HuggingFaceFW/finepdfs/data/{args.lang}/train/*.parquet"
 
     print("=" * 60)
@@ -145,7 +116,7 @@ def main():
         print("\nNo --output specified. Use --output to push to Hub.")
         return
 
-    # Rebuild query for streaming
+    # Rebuild query for streaming to parquet
     lf = (
         pl.scan_parquet(source)
         .filter(
@@ -168,14 +139,17 @@ def main():
     if args.limit:
         lf = lf.limit(args.limit)
 
-    print(f"\nStreaming to dataset (chunk_size={args.chunk_size})...")
-    ds = IterableDataset.from_generator(
-        lambda: polars_to_generator(lf, chunk_size=args.chunk_size),
-        features=FEATURES,
-    )
+    # Use sink_parquet for true streaming (minimal memory)
+    with tempfile.TemporaryDirectory() as tmpdir:
+        output_path = Path(tmpdir) / "data.parquet"
+        print("\nStreaming to parquet (sink_parquet)...")
+        lf.sink_parquet(output_path)
+
+        print(f"\nLoading parquet and pushing to {args.output}...")
+        ds = Dataset.from_parquet(str(output_path))
+        print(f"Dataset: {len(ds)} rows")
+        ds.push_to_hub(args.output, private=args.private)
 
-    print(f"\nPushing to {args.output} (streaming)...")
-    ds.push_to_hub(args.output, private=args.private)
     print(f"\nDone! https://huggingface.co/datasets/{args.output}")