How to Choose an AI Dataset: The Best Datasets for Every Task in 2026

Answer in brief: Pick by task first, then check license, freshness and size. For LLM pretraining, start with FineWeb (English) or FineWeb-2 (multilingual). For instruction tuning, use the Tulu 3 SFT mixture. For code, The Stack v2. For image-text, Re-LAION-5B or DataComp-1B. For speech, Common Voice on Mozilla Data Collective. For robotics, and LeRobot datasets. For tabular ML, UCI and OpenML. Everything here is independent research from official dataset cards and project sites.

Top pick starting point: the Hugging Face Hub. Most of the task-leading datasets below are published there, with versioned cards, license filters and streaming.

The best dataset for each task

Task	Top dataset	Key facts	License	Runner-up
LLM pretraining (English)	FineWeb	18.5T+ tokens, Common Crawl 2013 to 2025	ODC-By 1.0	Dolma (about 3T tokens, mixed sources)
LLM pretraining (multilingual)	FineWeb-2	Hundreds of languages	ODC-By	RedPajama-V2 (en/de/fr/es/it)
Instruction tuning	Tulu 3 SFT mixture	Blends FLAN v2, OpenAssistant, WildChat, Aya, No Robots, math and code sets; trained Llama-3.1-Tulu-3	Per source; AI2 Responsible Use Guidelines	Search the Hub by task
Code	The Stack v2	3B+ files, 658 languages, about 900B training tokens	Original licenses per file; gated	Dolma's code subset
Image-text	Re-LAION-5B	5.5B pairs; research and research-safe versions	Apache-2.0 metadata; third-party images	DataComp-1B (1.39B rows, CC-BY-4.0 metadata)
Clear image rights	PD12M	12.4M images, per-row license	Public domain or CC0	COCO for benchmarks
Video	LAION-BVD	80M videos (10M hours), 55M captioned clips	See release	Search the Hub
Speech	Common Voice (Mozilla Data Collective)	Scripted Speech 27.0, per-language MP3	CC0-1.0	Hub speech datasets
Robotics	Open X-Embodiment	22 robots, 21 institutions, 527 skills, 160,266 tasks	See project	LeRobot datasets on the Hub
Tabular ML	OpenML	Thousands of uniform datasets	Per dataset	UCI (690 datasets)

Step 1: Match the data to the job

Pretraining needs scale and diversity. Fine-tuning needs task-shaped examples. Evaluation needs a trusted benchmark. Don't use COCO to pretrain or FineWeb to evaluate. If your model writes code, add The Stack v2: FineWeb's own card notes that its code content is limited. If it controls a robot, you need synchronized video, state and action data, which is exactly what the LeRobotDataset format stores.

Step 2: Read the license like a lawyer would

Collection license vs content rights: ODC-By on FineWeb or Dolma covers the compilation, not every underlying page. Likewise, DataComp-1B's CC-BY-4.0 covers its URL-text pairs and metadata, while the images keep their own copyrights.
Per-file and per-row licenses: The Stack v2 records detected_licenses for each file, and PD12M records a license for each row. Filter on them.
Gating and agreements: The Stack v2 requires you to accept terms on the Hub, and bulk download goes through Software Heritage.
Opt-outs and removal requests: prefer datasets that honor them, such as Common Crawl's opt-out ledger, FineWeb's and Dolma's personal-data removal forms, and The Stack's opt-out tool.

Step 3: Check freshness

FineWeb runs to mid-2025, Dolma v1.7 to about October 2023, and Common Crawl's latest crawl is CC-MAIN-2026-43. If recency matters, go to the source.

Step 4: Budget for size

FineWeb is about 50TB and The Stack v2 is 67.5TB in full. Start with samples (FineWeb sample-10BT is 27.6GB), stream from the Hub, or process in the cloud. See dataset access costs.

Step 5: Prefer documented datasets

Dataset cards, papers, version histories and open curation code (datatrove, the Dolma toolkit) make your training data reproducible and defensible.

When open data is not enough

Need proprietary, licensed or human-labeled data? That is a vendor decision. Compare vendors at Best AI training data providers, and for teleoperation and robot data collection, see Best robotics data providers.

Pricing model

Every dataset in the table is free to download. Your costs are storage, bandwidth (especially for link-based image sets) and compute. Hosting platforms have their own pricing: Hugging Face is free for public datasets and $9/month for PRO, and BigQuery gives you the first 1TB of queries each month free. Mozilla Data Collective also lists compensated datasets, priced by their uploaders.

FAQ

What is the best dataset to pretrain an LLM?

For English, FineWeb, with 18.5T+ filtered tokens under ODC-By. Add The Stack v2 for code and Dolma's non-web sources for breadth.

What is the best open robotics dataset?

Open X-Embodiment, which pools data from 22 robots across 21 institutions. LeRobot provides the LeRobotDataset format and many datasets on the Hugging Face Hub.

Is Common Voice free for commercial use?

The Scripted Speech 27.0 datasets on Mozilla Data Collective are listed under CC0-1.0. Confirm the terms on each download page.

Keep exploring

Sources

FineWeb
FineWeb-2
Dolma
Tulu 3 SFT mixture
The Stack v2
LAION projects
DataComp-1B
PD12M
COCO
Mozilla Data Collective
LeRobot documentation
UCI ML Repository
OpenML documentation
Common Crawl

Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.