FineWeb-2 is a large-scale, high-quality web text corpus derived from Common Crawl, built through extensive filtering, deduplication, and language identification steps to support the training of modern large language models. It emphasizes text quality, diversity, and transparency, and includes rich crawl-level metadata inherited from Common Crawl.

This release contains a 75,000-entry sample drawn from FineWeb-2 and enriched with additional symbolic metadata layers aimed at improving interpretability, licensing awareness, and semantic analysis of web text.

Language Distribution

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

Language	Subset	Number of Entries	Percentage
English	en	15,000	20%
Italian	it	15,096	20%
French	fr	15,054	20%
German	de	15,055	20%
Spanish	es	15,043	20%

Sampling Strategy

This dataset represents a small sample of the original FineWeb-2 corpus. The sampling preserves crawl-level provenance and the original FineWeb-2 filtering guarantees.

The goal is not to replace FineWeb-2, but to provide a research-oriented subset enriched with structured metadata that can support downstream analyses of licensing, content characteristics, and semantic properties.

Metadata Enrichment

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

License-Related Metadata

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

whether the original URL is still reachable,
whether the website’s robots.txt allows crawler access,
whether the website explicitly mentions a license governing its content.

These signals are intended to support large-scale auditing and filtering, not to provide definitive legal judgments.

Semantics-Related Metadata

Semantics-related metadata is designed to provide a structured, symbolic interpretation of the text. Four main areas are covered:

Formal features Including text length, readability, informativeness, and structural complexity.
Entity focus Detection of people, organizations, temporal references, and geographical locations.
Domain and topic classification Assignment of domains and subdomains describing the subject matter of the text.
Content risk indicators Detection of potentially biased content, sensitive information, opinions, negativity, and personal data.

All semantics-related metadata has been extracted using an entirely symbolic processing stack, chosen for cost efficiency, scalability, and reproducibility.

In particular, semantic annotations were generated using the proprietary expert.ai knowledge graph.

Methodology Notes

No neural models were used for metadata extraction.
All annotations are deterministic and rule-based.
This design prioritizes computational efficiency, interpretability, and ease of large-scale deployment.

Work in Progress

This dataset represents an initial sample release. Work is ongoing to extend the same metadata enrichment pipeline to the entirety of FineWeb-2.

All future steps, methodological details, and validation results will be documented in a dedicated technical report.

Data Fields

Each entry includes the following fields:

Field	Type	Description
`text`	string	Main text content
`id`	string	Original unique identifier from Common Crawl
`dump`	string	Common Crawl dump this sample was extracted from
`url`	string	URL of the original web page
`date`	string	Crawl date (from Common Crawl)
`file_path`	string	S3 path of the Common Crawl WARC file
`language`	string	ISO 639-3 language code
`language_score`	float64	Language prediction score from GlotLID
`language_script`	string	Script of the text (e.g. Latn)
`minhash_cluster_size`	int64	Size of the MinHash cluster
`top_langs`	string	Language-script pairs with classifier scores
`eai_lenChar`	string	Length of the text in characters
`eai_lenClass`	string	Length class (clause, paragraph, article, book, etc.)
`eai_readabilityLIX`	string	LIX readability index
`eai_readabilityEAI`	string	Structural and semantic readability index
`eai_readabilityCScore`	string	Informativeness score
`eai_quality`	string	Indicators of processing difficulty
`eai_geoClass`	string	Detected geographic areas
`eai_timeRef`	string	Detected temporal references
`eai_categories`	string	Topics and domains
`eai_subject`	string	Types of referents detected
`eai_metrics`	string	Relevance of detected entity types
`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 (EU terminology) identifiers
`eai_bias`	string	Presence of biased content
`eai_sensitiveContent`	string	Presence of sensitive content
`eai_opinions`	string	Opinionated language indicators
`eai_negativity`	string	Negative sentiment
`eai_privacy`	string	Presence of personal data
`eai_source`	string	Source web domain
`eai_urlReachable`	string	Whether the URL is currently reachable
`eai_robotsTxt`	string	Robots.txt crawler permissions
`eai_licenseInfo`	string	License information detected on the website

Intended Use

This dataset is intended for:

research on licensing-aware data selection,
semantic analysis of large-scale web corpora,
multilingual text quality assessment,
development and evaluation of dataset auditing pipelines.

It is not intended to replace FineWeb-2, but to complement it with richer symbolic metadata.

Licensing

This dataset inherits the licensing considerations of the original FineWeb-2 corpus. The additional metadata is provided as-is and does not constitute legal advice regarding the usability of the underlying content.

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

Downloads last month: 1

You need to agree to share your contact information to access this dataset

fineweb-2-sample-75k-meta

Overview