FineWeb 2
The FineWeb quality-filtering methodology extended to over 1,000 languages (1,868 language-script pairs) — roughly 8TB of compressed, deduplicated web text (~3 trillion words) from 96 CommonCrawl snapshots. Per-language pipelines were tuned and validated so filtered data beats unfiltered in ablation training. The default multilingual pretraining corpus for open models.
from datasets import load_dataset
ds = load_dataset("HuggingFaceFW/fineweb-2")Preview a sample row
{
"text": "Fotowoltaika w Polsce rozwija się dynamicznie od 2019 roku...",
"id": "<urn:uuid:...>",
"dump": "CC-MAIN-2024-18",
"language": "pol",
"language_script": "Latn",
"language_score": 0.99
}Frequently asked questions
Can I use FineWeb 2 commercially?
Yes — FineWeb 2 is released under ODC-By 1.0, a permissive license that allows commercial use, including training models you ship in a product. Check the dataset card for attribution requirements before release.
How much data does FineWeb 2 contain, and do I need all of it?
FineWeb 2 contains 8TB Compressed. You rarely need all of it: for style and format fine-tuning, a few hundred to a few thousand examples are enough — load a slice (e.g. split="train[:1000]") and scale up only if quality plateaus.
What is FineWeb 2 best used for?
Pretraining or continued pretraining in languages other than English. It belongs to the Pretraining section of our dataset hub, where you'll find alternatives and complementary sets.