Ploba logo

Find the right AI tool for the job: apps, agents, APIs, MCP servers and datasets, explained plainly.

Get the AI Digest

Discover

  • Explore
  • Top Charts
  • Categories
  • Agents
  • MCP servers
  • Developer tools

Learn & community

  • Learn AI
  • AI Glossary
  • Blog
  • Hackathons
  • People
  • AI Digest

For makers

  • List your tool
  • Docs
  • Claim a listing
  • Advertise
  • Write for us

Company

  • About
  • Careers
  • Changelog
  • Trust
  • Security
  • Contact

© 2026 Ploba. All rights reserved.

  • Privacy
  • Terms
  • Cookies
  • Advertising
  • All policies
llms.txt
Ploba logo
DiscoverHackathonsPeopleLearnAI Digest
Sign inGet started
HomeTop ChartsExploreCategories
TypesWeb apps1,894Agents585Plugins416APIs376Mobile apps152Datasets85
AdvertisingYour product hereReach people browsing AI tools on Ploba.Advertise hereAdvertisingYour product hereReach people browsing AI tools on Ploba.Advertise here
  1. Explore
  2. /
  3. AI
  4. /
  5. Ultra-FineWeb
AI
Ul

Ultra-FineWeb

by OpenBMB

Ultra-FineWeb is OpenBMB's filtered web pre-training dataset, with about 1T English tokens and 120B Chinese tokens. It is built by applying a verification-based quality filter to FineWeb and Chinese FineWeb data.

Dataset·Free

View on Hugging Face(opens in a new tab)

About Ultra-FineWeb

Key features of Ultra-FineWeb

  • About 1T English tokens and 120B Chinese tokens
  • Built from FineWeb and Chinese FineWeb sources
  • Verification-based filtering pipeline with a lightweight fastText classifier
  • L1 release with cleaning, heuristic filtering and deduplication
  • L3 release of rewritten and Q&A-style synthetic data
  • Common Crawl snapshots covered up to CC-MAIN-2025-51 in the L1 release
  • Technical report on arXiv

Who Ultra-FineWeb is for

  • Teams pre-training large language models
  • Researchers studying data filtering and quality
  • Builders of Chinese and English training corpora

How to use Ultra-FineWeb

  1. 1

    Step 1: Open the dataset page on Hugging Face and read the card to choose between the English, Chinese, L1 and L3 variants.

  2. 2

    Step 2: Check the license of each underlying source dataset for your intended use.

  3. 3

    Step 3: Download or stream the files you need with the Hugging Face datasets library or hub tools.

  4. 4

    Step 4: Optionally use the separate Ultra-FineWeb classifier to filter your own web text.

  5. 5

    Step 5: Mix the data into your pre-training run and evaluate on your own benchmarks.

Ultra-FineWeb FAQ

How large is Ultra-FineWeb?

The card lists about 1T English tokens and about 120B Chinese tokens. The L3 release adds more than 400B English and 200B Chinese tokens.

What license does Ultra-FineWeb use?

The project is released under Apache 2.0. The card also says to check the license of each dataset it was built from.

What is the difference between L1 and the main dataset?

L1 is the data after basic cleaning, heuristic filtering and deduplication. The main Ultra-FineWeb set is selected from it with the Ultra-FineWeb classifier.

Is the filtering classifier available?

Yes. The classifier is published as a separate Hugging Face repository and is built on fastText.

Details

Pricing
FreeReleased under the Apache 2.0 license. Because the data is built from several source datasets, the card tells users to check each source dataset's license.
Type
Dataset
Categories
AIDataResearch
Added
Sep 17, 2026
OpPublisherOpenBMB

Built from public sources

We put this page together from Ultra-FineWeb’s own website and docs so you can compare it with other tools. Ploba isn’t affiliated with Ultra-FineWeb.

Work at Ultra-FineWeb? Claim the listing to manage it yourself.

Something wrong here?

More from OpenBMB

View publisher
USUltraData SFT 2605OpenBMB · FreeUltraData-SFT-2605 is OpenBMB's set of over 15 million supervised fine-tuning samples used to post-train a small on-device language model. It covers math, code, knowledge and instruction following, with both deep-thinking and non-thinking data.AI · FreeUMUltraData MathOpenBMB · FreeUltraData-Math is an OpenBMB math pre-training dataset of over 290B tokens in three quality tiers, built to improve mathematical reasoning in language models.AI · FreeUSUltraData SFT Agent 2609OpenBMB · FreeUltraData-SFT-Agent-2609 is OpenBMB's dataset of 483,661 agent trajectories for instruction tuning, covering tool use, search, code and general agent tasks. It keeps full trajectories with tool calls, observations and errors.Developer Tools · Free

Related tools

See more
AWABot World Explorer 500hACVLab · FreeABot World Explorer 500h is a dataset of 30,969 action-conditioned video episodes from ACVLab. Each episode pairs an MP4 with keyboard actions, captions and a COLMAP camera model.AI · FreeX1Xperience 10MRopedia AI · FreeXperience-10M is Ropedia AI's egocentric multimodal dataset of 10 million experiences and 10,000 hours of first-person recordings. It is for work on embodied AI, robotics and world models.AI · FreeEgEgoStandardLightwheel AI · FreeEgoStandard is Lightwheel's egocentric video dataset pairing head-view video with synchronized 3D hand pose. It is the 90,000-hour head-view line of EgoSuite-Open100K.AI · Free