storagestuffnobody/finepdf-subsampled

FinePDFs subsampled

Publication is complete. All output counts match the previously audited selection.

This is the previously selected approximately 150B-token English subset of HuggingFaceFW/finepdfs, using source revision 220bac3acbf07789502c621d2d33952f51ac7f86. Selection is deterministic: every qualifying document is retained, without an additional random sample.

Split	Documents	Stored tokens
train	8,780,374	149,232,518,361
test	20	405,303
total	8,780,394	149,232,923,664

Selection

All conditions must hold:

• full_doc_lid == "eng_Latn"
• is_truncated == false
• token_count > 8192
• ocr_quality_scores is nonempty and nonnull; every score is finite, nonnull and strictly greater than 1.75.

The length and OCR cutoffs are strict. No page-average language-score filter was added. Tokens are whole-document sums of the stored upstream token_count; the text was not retokenized. All 23 original fields are retained, including full text, document ID, source URL, crawl metadata, page boundaries, language labels and quality scores. Dictionary storage encoding is normalized to strings without changing field values. There is no new text cleaning, truncation, chunking or deduplication. Original train/test membership is preserved; the small upstream test split remains separate.

Use

from datasets import load_dataset

train = load_dataset("storagestuffnobody/finepdf-subsampled", split="train", streaming=True)
document = next(iter(train))
text = document["text"]

Parquet files can also be read directly with PyArrow. source_manifest.json identifies every source file, its upstream SHA256 and expected selected counts. Each source folder contains a manifest.json with output file hashes and count verification. The final manifest.json and checksums.sha256 index all output shards. Source source_mtime_ns values are local inventory observations, not a requirement on other users' downloads.

Interpretation and provenance

OCR quality scores evaluate sampled text. Passing this screen does not prove every page was recovered or that a document is free of layout errors or OCR hallucinations. The upstream truncation flag concerns the source crawl artifact; it does not certify extraction completeness. Metadata language labels are model predictions.

The source dataset was created by Hynek Kydlíček, Guilherme Penedo and Leandro von Werra / Hugging Face. This subset retains the upstream ODC-By 1.0 attribution and source dataset documentation, including its reference to CommonCrawl terms. This repository contains a filtered copy of extracted text, not newly extracted PDF files.