Free🇺🇸
Hugging Face pretraining dataset of 1.3 trillion tokens of educational web pages, selected from Common Crawl by an AI quality classifier.
FineWeb2 overview and key facts (October 2026)
Multilingual pretraining dataset from Hugging Face covering more than 1,000 languages, including a large German subset.
Free
What is FineWeb2?
FineWeb2 is an open multilingual pretraining dataset built from Common Crawl, with separate subsets for more than 1,000 languages. It gives teams training data for languages other than English.
FineWeb2 was built from 96 Common Crawl snapshots from summer 2013 to April 2024 and processed with Hugging Face's datatrove library. After deduplication and filtering it holds about 20 terabytes, 5 billion documents and more than 3 trillion words. Each language is a separate subset, so teams can download only what they need. The processing pipeline is published on GitHub and the dataset is released under ODC-By 1.0.
FineWeb2 AI features and use cases
- Trains models for German and other languagesTeams that build or adapt language models for German can download the German subset alone. It gives them hundreds of billions of words of cleaned web text without running their own crawl.
- Supports continued pretrainingAn existing open model can be trained further on one language subset to improve fluency and vocabulary in that language. This is a common step before fine-tuning a model for company tasks in German.
- Reproduces the pipeline on own dataBecause the processing pipeline is public, data teams can apply the same language identification, deduplication and filters to their own crawls and compare results. This makes it easier to explain how training data was prepared, for example in model documentation.
What sets FineWeb2 apart
- One subset per language, including a German subset of about 1.5 terabytes of raw text.
- Fully reproducible pipeline published on GitHub.
Where FineWeb2 is strongest
United States Available internationally
FineWeb2 is multilingual and available worldwide; its German subset holds about 262 billion words from 496 million documents.
Questions to ask before you buy FineWeb2
Only emails and IP addresses are anonymized; other personal data and copyrighted text from web pages can remain. ODC-By requires attribution, and Common Crawl's terms also apply. The card describes an opt-out process for personal and sensitive information. Plan storage for large language subsets.
FineWeb2 reviews
No reviews yet. Be the first to share how your team uses it.
Every review shows its verification level. We read each one before it goes live. How we check reviews
Questions about FineWeb2
How much German data does FineWeb2 contain?
The German subset deu_Latn holds about 262 billion words from 496 million documents, about 1.5 terabytes of text before compression.
Can we download a single language?
Yes. Each language is a separate subset and can be loaded on its own with the datasets library, huggingface_hub or datatrove.
Which license applies?
FineWeb2 is released under the Open Data Commons Attribution License 1.0. Commercial use is allowed with attribution, subject to Common Crawl's terms.
FineWeb2 update historyno updates yet
No changes recorded for FineWeb2 yet.
Changes the vendor made and confirmed user reports, each checked against the vendor's own pages. We review every tool monthly.
Alternatives to FineWeb2
Tools in the same category (Training data) with the most overlapping markets.
Free🇺🇸
Instruction-tuning mix of 939,344 examples from the Allen Institute for AI, used to train the Tulu 3 models.
Free🇫🇷
Open pretraining corpus of 2.27 trillion tokens from Pleias in Paris, containing only uncopyrighted or freely licensed documents.
Contact for pricing · Free planAvailable internationally
Free, open-source prompt library, formerly Awesome ChatGPT Prompts, available as a website, MCP server or self-hosted company library.