| """ |
| Inspect locally-downloaded FineWeb2-HQ data. |
| |
| The dataset is already on disk as parquet shards under: |
| |
| <data_root>/<language>/*.parquet |
| |
| HuggingFace `datasets` (and `pyarrow`) read parquet directly, so there is no |
| need to re-download or transcode to JSONL. This script only verifies that the |
| shards exist and reports per-language statistics (shard count, row count, |
| text-length summary, optional sample preview). |
| |
| Typical use: |
| |
| python prepare_data_local.py \ |
| --data-root /root/userdata/liangyan/routing_analysis/datasets/FineWeb2-HQ \ |
| --languages dan_Latn fas_Arab swe_Latn \ |
| --preview |
| """ |
|
|
| import argparse |
| from pathlib import Path |
|
|
| import pyarrow.parquet as pq |
|
|
|
|
| def list_shards(lang_dir: Path) -> list[Path]: |
| """Return the parquet shards for a language, sorted by filename.""" |
| return sorted(lang_dir.glob("*.parquet")) |
|
|
|
|
| def summarize_language( |
| lang: str, |
| lang_dir: Path, |
| sample_rows: int = 100, |
| preview: bool = False, |
| ): |
| """Print shard count, row count, text-length stats, and optional preview.""" |
| print(f"\n{'=' * 60}") |
| print(f"Language: {lang}") |
| print(f"Path: {lang_dir}") |
| print(f"{'=' * 60}") |
|
|
| if not lang_dir.is_dir(): |
| print(" [missing] directory does not exist") |
| return |
|
|
| shards = list_shards(lang_dir) |
| if not shards: |
| print(f" [missing] no .parquet files in {lang_dir}") |
| return |
|
|
| |
| total_bytes = sum(p.stat().st_size for p in shards) |
| total_rows = 0 |
| schema = None |
| for shard in shards: |
| meta = pq.read_metadata(shard) |
| total_rows += meta.num_rows |
| if schema is None: |
| schema = pq.read_schema(shard) |
|
|
| print(f" Shards: {len(shards)}") |
| print(f" Rows: {total_rows:,}") |
| print(f" Size: {total_bytes / 1024 / 1024 / 1024:.2f} GB") |
| print(f" Columns: {', '.join(schema.names)}") |
|
|
| |
| first_shard = shards[0] |
| table = pq.read_table(first_shard, columns=["text"]) |
| sample = table.slice(0, sample_rows).column("text").to_pylist() |
| lengths = [len(t) for t in sample if t] |
| if lengths: |
| avg_len = sum(lengths) / len(lengths) |
| print( |
| f" Text len: avg={avg_len:.0f} " |
| f"min={min(lengths)} max={max(lengths)} (first {len(lengths)} rows)" |
| ) |
|
|
| if preview and sample: |
| snippet = sample[0] or "" |
| print("\n --- preview (row 0, first 300 chars) ---") |
| print(" " + snippet[:300].replace("\n", "\n ")) |
|
|
|
|
| def main(): |
| parser = argparse.ArgumentParser( |
| description="Inspect locally-downloaded FineWeb2-HQ parquet data." |
| ) |
| parser.add_argument( |
| "--data-root", |
| type=Path, |
| default=Path( |
| "/root/userdata/liangyan/routing_analysis/datasets/FineWeb2-HQ" |
| ), |
| help="Directory containing one subdirectory per language.", |
| ) |
| parser.add_argument( |
| "--languages", |
| nargs="+", |
| default=["dan_Latn", "fas_Arab", "swe_Latn"], |
| help="Language codes (subdirectory names) to inspect.", |
| ) |
| parser.add_argument( |
| "--sample-rows", |
| type=int, |
| default=100, |
| help="Rows to load from the first shard for text-length stats.", |
| ) |
| parser.add_argument( |
| "--preview", |
| action="store_true", |
| help="Print a short text snippet from the first row of each language.", |
| ) |
| args = parser.parse_args() |
|
|
| print("FineWeb2-HQ local data inspection") |
| print(f" data root: {args.data_root}") |
| print(f" languages: {', '.join(args.languages)}") |
|
|
| for lang in args.languages: |
| summarize_language( |
| lang=lang, |
| lang_dir=args.data_root / lang, |
| sample_rows=args.sample_rows, |
| preview=args.preview, |
| ) |
|
|
| print( |
| "\nDone. To load a language during training, point HuggingFace " |
| "`datasets` at the parquet shards directly, e.g.:\n" |
| " from datasets import load_dataset\n" |
| " ds = load_dataset(\n" |
| " 'parquet',\n" |
| f" data_files=f'{args.data_root}/<lang>/*.parquet',\n" |
| " split='train',\n" |
| " streaming=True,\n" |
| " )" |
| ) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
