Dataset catalog
Filter by type, access, and pricing. Specs show before you open the product page.
tau-bench (τ-bench)
Tool-agent-user benchmark of 165 customer-service tasks (retail 115, airline 50) with policy-following and pass@k evaluation.
WebArena
Realistic self-hosted web environment: 812 long-horizon tasks over shopping, forum, GitLab, CMS and maps with execution-based evaluation.
MiniWoB++
Farama-maintained suite of 128 synthetic mini web-interaction tasks (button/form/date/inbox) with DOM+screenshot observations.
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.
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.
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.
PettingZoo (Farama)
The standard Python API and reference environment suite for multi-agent RL, covering Atari (multiplayer), Classic board/card games, MPE, SISL, and Butterfly.
Procgen Benchmark
OpenAI's Procgen Benchmark is a suite of 16 procedurally-generated, Atari-like Gym/Gym3 reinforcement-learning environments built to measure sample efficiency and generalization in RL.
BabyAI
Grid-world instruction-following platform with a compositional synthetic 'Baby Language'; 19 levels of increasing difficulty, now part of Minigrid.
Jericho
Open-source Python RL environment from Microsoft Research that connects agents to 57 human-made interactive-fiction text games via a modified Frotz Z-machine interpreter, with reward defined as in-game score deltas.
Overcooked-AI
A two-agent gridworld cooking environment for benchmarking human-AI and multi-agent coordination, where agents cooperatively prepare and deliver soups for a shared sparse reward.
TextWorld
Microsoft's sandbox engine for procedurally generating and playing text-adventure games to train and evaluate RL agents; underlies ALFWorld.