Dataset catalog
Filter by type, access, and pricing. Specs show before you open the product page.
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.
Stanford Human Preferences (SHP)
385,563 collective human preference pairs mined from 18 Reddit Q&A subreddits, where the more-helpful of two comments is inferred from Reddit scores and timestamps.
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.
MiniWoB++
Farama-maintained suite of 128 synthetic mini web-interaction tasks (button/form/date/inbox) with DOM+screenshot observations.
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.
WebGPT Comparisons
19,578 human-scored answer-pair comparisons from OpenAI's WebGPT project, used to train the WebGPT long-form question-answering reward model.
ScienceWorld
Text-based interactive environment covering elementary-science curriculum tasks (30 task types) with a dense reward in [0,1].
WebShop
Simulated e-commerce shopping environment: 1.18M real products and 12,087 instructions with a search/click action space and dense reward.
CALVIN
CALVIN is an open-source simulated benchmark of ~24 hours of human teleoperated play data (6 h in each of 4 environments A/B/C/D) with crowd-sourced natural-language annotations for learning long-horizon, language-conditioned robot manipulation over 34 tasks.
Anthropic HH-RLHF
Anthropic's human preference dataset of chosen/rejected assistant dialogue pairs for training helpful and harmless (HH) reward models via RLHF.
GSM8K
GSM8K is a dataset of 8.5K human-written grade-school math word problems, each paired with a multi-step natural-language solution that ends in a single final numeric answer marked by '####'.
Crafter
Crafter is an open-world 2D survival reinforcement-learning environment where agents are evaluated by the 22 semantically meaningful achievements they can unlock in each procedurally generated episode.
MiniHack
MiniHack is an open-source sandbox framework, built on the NetHack Learning Environment, for designing and running procedurally generated NetHack-based reinforcement-learning tasks (navigation and skill-acquisition), exposed as pre-registered Gymnasium environments.
ALFWorld
Text-based embodied household RL environment aligning ALFRED tasks with TextWorld; 134 unseen eval games across 6 task types with an optional vision variant.
StrategyQA
A yes/no question-answering benchmark whose questions require implicit multi-step (strategy) reasoning, graded by boolean exact-match against a gold true/false answer.
NetHack Learning Environment (NLE)
A Gymnasium reinforcement-learning environment that wraps the roguelike game NetHack 3.6.7, exposing symbolic game-state observations and a set of programmatically-rewarded goal tasks.