Dataset catalog
Filter by type, access, and pricing. Specs show before you open the product page.
UltraChat 200k
HuggingFaceH4's heavily filtered subset of UltraChat (~208k train_sft dialogues) used to train the Zephyr chat models.
MetaMathQA
395k math QA pairs bootstrapped from GSM8K and MATH via rephrasing, self-verification and backward (FOBAR) augmentation; text solutions only.
MathInstruct
262k math instruction examples combining chain-of-thought and program-of-thought (executable Python) rationales compiled from 13 datasets.
Evol-CodeAlpaca v1
111K English code instruction/output pairs — an Evol-Instruct augmentation of CodeAlpaca-20k using GPT-4 across 10 evolution strategies.
Dolphin
Open Orca-style FLAN reproduction: ~892K FLANv2 completions from GPT-4 plus ~2.84M from GPT-3.5, filtered to remove refusals/alignment.
Evol-Instruct-Code-80k-v1
78K code instruction/output pairs — an open reproduction of WizardCoder's Evol-Instruct-Code (CodeAlpaca run through 3 evolution rounds).
OpenOrca
An open collection of ~2.94M FLAN-Collection instructions paired with GPT-4/GPT-3.5 'Orca-style' reasoning-trace responses for supervised instruction fine-tuning.
WizardLM Evol-Instruct V2 196k
Complexity-evolved (Evol-Instruct) instruction conversations derived from Alpaca/ShareGPT for WizardLM SFT; HF split ships 143K rows.
Self-Instruct
Machine-generated instruction-following dataset created by bootstrapping GPT-3 from 175 seed tasks, released as ~82K prompt/completion demonstrations for instruction tuning.
LIMA
GAIR's 1,030-example gated dataset of high-quality curated prompts and responses used in the 'Less Is More for Alignment' study.
OpenAssistant oasst1
Crowdsourced multilingual assistant conversation trees with human quality ratings; HF splits are 84.4k train / 4.4k validation messages.
Databricks Dolly 15k
~15k human-authored instruction/response records across 8 task categories, written by Databricks employees for commercial-friendly instruction tuning.
CodeAlpaca 20K
20K single-function code instructions (instruction/input/output) generated via Self-Instruct with text-davinci-003 from 21 seed tasks.
Stanford Alpaca (and Alpaca-Cleaned)
52K single-turn English instruction-following examples generated from OpenAI text-davinci-003 via Self-Instruct; cleaned mirror fixes ~51.8K rows.
FLAN Collection (Flan v2)
Community re-release of Google's Flan 2022 Collection: 1,836 tasks across Flan/T0/NIV2/CoT/Dialog in zero/few-shot × option/no-option formats.
Super-NaturalInstructions
A benchmark of 1,616 diverse NLP tasks, each paired with an expert-written declarative instruction (Definition) plus positive/negative demonstration examples and many input/output instances, used for instruction-tuning and cross-task generalization.
Unnatural Instructions
A large instruction-tuning dataset of ~68K instruction-input-output triplets (240K with paraphrases) generated almost entirely by OpenAI's text-davinci-002 from three seed examples, with virtually no human labor.
NuminaMath-1.5
~896k competition-math problems (Chinese high-school to IMO level) with chain-of-thought solutions; improved successor to NuminaMath-CoT.
OpenR1-Math-220k
220k competition math problems, each with 2-4 DeepSeek-R1 chain-of-thought traces verified by Math-Verify and Llama-3.3-70B; text reasoning, no code.