README.md

metadata

tags:
  - statistics
  - polars
  - finepdfs-edu
license: odc-by

Statistics for HuggingFaceFW/finepdfs-edu (Dutch (Latin script))

Aggregate statistics computed using Polars streaming on the HuggingFaceFW/finepdfs-edu dataset.

Performance

Processed 847,361 documents in 125.88 seconds.

Step	Time
Global stats	34.19s
Language stats	31.81s
Extractor stats	29.46s
Dump stats	30.42s
Total	125.88s

Speed comes from Polars only reading metadata columns (not the text column), thanks to Parquet's columnar format and lazy evaluation.

How This Was Generated

This dataset demonstrates Polars streaming aggregation with HuggingFace Hub integration. Thanks to polars#25521, scan_parquet with hf:// paths now uses far fewer API calls (139 → 1 for finepdfs-edu).

uv run polars-hf-stats.py --lang nld_Latn --output-repo davanstrien/finepdfs-nld-stats

Global Summary

Metric	Value
Language	Dutch (Latin script) (nld_Latn)
Total Documents	847,361
Total Tokens	5,424,871,452
Average Tokens/Doc	6,402
Truncated Documents	207,376 (24.5%)
Unique Languages	1
Unique Extractors	2
Unique Dumps	106

Configs

• global_stats - Overall summary (1 row)
• language_stats - Per-language aggregations
• extractor_stats - Per-extractor aggregations
• dump_stats - Per-dump aggregations

Usage

from datasets import load_dataset

# Load all configs
global_stats = load_dataset("davanstrien/finepdfs-nld-stats", "global_stats")
lang_stats = load_dataset("davanstrien/finepdfs-nld-stats", "language_stats")
extractor_stats = load_dataset("davanstrien/finepdfs-nld-stats", "extractor_stats")
dump_stats = load_dataset("davanstrien/finepdfs-nld-stats", "dump_stats")

Source

• Dataset: HuggingFaceFW/finepdfs-edu
• Language: nld_Latn
• Script: polars-hf-stats.py