FinePDFs is a large-scale collection of document-level texts extracted from PDF files, sourced primarily from Common Crawl. The dataset emphasizes high-quality document extraction, structural coherence, and large-scale coverage of technical, scientific, educational, and administrative content commonly distributed in PDF form.

This release contains a 75,000-document sample drawn from FinePDFs and enriched with additional symbolic metadata layers designed to improve interpretability, licensing awareness, and semantic analysis of document-based web content.

Language Distribution

The dataset includes five languages, with an equal number of documents per language:

Language	Subset	Number of Entries	Percentage
English	en	15,000	20%
Italian	it	15,009	20%
French	fr	15,006	20%
German	de	15,003	20%
Spanish	es	14,997	20%

Sampling Strategy

This dataset represents a sample of the original FinePDFs corpus.
The sampling preserves language diversity and original FinePDFs filtering and extraction guarantees.

The goal is not to replace FinePDFs, but to provide a research-oriented subset enriched with structured metadata that enables deeper analysis of document properties, licensing signals, and semantic content.

Metadata Enrichment

Each document has been enriched with two complementary categories of metadata.

License-Related Metadata

License-related metadata aims to assess whether the source documents may be considered permissively usable. For each document, we check:

whether the original source URL is still reachable,
whether the hosting domain’s robots.txt allows crawler access,
whether the website mentions any license governing the document content.

These indicators are intended to support large-scale dataset auditing and filtering rather than provide definitive legal conclusions.

Semantics-Related Metadata

Semantics-related metadata provides a symbolic, structured characterization of document content. Four main areas are explored:

Formal document features
Including length, structural complexity, readability, and informativeness.
Entity-centric analysis
Detection of people, organizations, temporal references, and geographical locations.
Domain and topic classification
Assignment of document domains and subdomains reflecting subject matter.
Content risk and sensitivity indicators
Identification of biased language, sensitive information, opinionated content, negative sentiment, and personal data.

All semantic annotations were extracted using an entirely symbolic processing pipeline, selected for scalability, cost efficiency, and reproducibility.

In particular, semantics-related metadata was derived using the proprietary expert.ai knowledge graph.

Methodology Notes

No neural models were used for metadata extraction.
All enrichment steps are deterministic and rule-based.
The pipeline is designed for efficient processing of large-scale document collections.

Work in Progress

This dataset represents an initial public sample of an ongoing enrichment effort.
Future releases aim to extend the same metadata extraction pipeline to the entirety of the FinePDFs corpus.

A comprehensive technical report will document methodology, coverage, and validation results.

Data Fields

Each entry contains the following fields:

Field	Type	Description
`text`	string	Main text content
`id`	string	Unique identifier for this sample
`dump`	string	Common Crawl dump this sample was part of
`url`	string	URL of the original page where the text was present
`date`	string	Crawl date (from Common Crawl)
`file_path`	string	S3 path for the individual Common Crawl WARC file containing this sample
`offset`	int	Offset in the Common Crawl WARC file containing this sample
`language`	string	ISO 639-3 code for the language and script of this sample
`per_page_languages`	list[string]	Per-page ISO 639-3 language and script codes
`page_average_lid`	string	ISO 639-3 language and script detected by averaging LID scores across pages
`page_average_lid_score`	float	Score of the top-detected language from page-level averaging
`full_doc_lid`	string	ISO 639-3 language and script detected from the first 40k characters
`full_doc_lid_score`	string	Score of the top-detected language from full-document LID
`is_truncated`	bool	Indicates whether the document is truncated in Common Crawl
`extractor`	string	PDF extractor used for this sample (`docling` or `rolmOCR`)
`page_ends`	list[int]	Indices denoting the end position of each page (exclusive)
`token_count`	int	Number of tokens computed using the GPT-2 tokenizer
`eai_lenChar`	string	Document length in characters
`eai_lenClass`	string	Length-based document class
`eai_readabilityLIX`	string	LIX readability index
`eai_readabilityEAI`	string	Structural and semantic readability score
`eai_readabilityCScore`	string	Informativeness classification
`eai_quality`	string	Indicators of text quality issues
`eai_geoClass`	string	Detected geographic references
`eai_timeRef`	string	Detected temporal references
`eai_categories`	string	Domain and topic labels
`eai_subject`	string	Referent types present
`eai_metrics`	string	Entity relevance metrics
`eai_idGeonames`	string	GeoNames identifiers
`eai_idWikiPeople`	string	Wikipedia people identifiers
`eai_idWikiOrg`	string	Wikipedia organization identifiers
`eai_idWikiGeo`	string	Wikipedia location identifiers
`eai_idWikiOther`	string	Other Wikipedia entity identifiers
`eai_idGoogleKGraph`	string	Google Knowledge Graph identifiers
`eai_idIATE`	string	IATE terminology identifiers
`eai_bias`	string	Indicators of biased content
`eai_sensitiveContent`	string	Sensitive content flags
`eai_opinions`	string	Opinionated language indicators
`eai_negativity`	string	Negative sentiment
`eai_privacy`	string	Presence of personal data
`eai_source`	string	Source domain
`eai_urlReachable`	string	URL reachability status
`eai_robotsTxt`	string	Robots.txt permissions
`eai_licenseInfo`	string	License information detected

Intended Use

This dataset supports:

analysis of document-level web corpora,
research on licensing-aware dataset construction,
multilingual document quality assessment,
development of symbolic and hybrid data auditing tools.

It complements, rather than replaces, the original FinePDFs dataset.

Licensing

This dataset inherits the licensing considerations of the original FinePDFs corpus.
The additional metadata is provided as-is and does not constitute legal advice.

Users are responsible for ensuring compliance with applicable licenses and usage terms.

Downloads last month: 9