DatoricDataset Catalog
Training data for frontier models.
Rights-cleared multimodal datasets spanning computer use, voice, egocentric video, human manipulation, robotics, and gameplay. Each ships with full annotation schemas, quality metrics, and QA review. Gated mirrors with sample data are available on Hugging Face.
Agentic
Computer-Use Traces Dataset
250,000 traces · 15,000 hours
250,000 global computer-use traces with screen states, browser sessions, UI actions, task instructions, and completion outcomes for training AI agents, computer-use models, and workflow automation systems.
View datasetVoice
TTS Voice Dataset
20,000 hours · 48kHz · 30+ languages
20,000-hour 48kHz global conversational audio dataset across popular and underrepresented languages. Built for TTS, voice cloning, speech synthesis, speaker modeling, and multilingual speech AI.
View datasetPhysical AI
Egocentric Residential Video Dataset
100,000 hours · 1080p+ · first person
100,000-hour 1080p+ egocentric video dataset of residential chores recorded in real homes, with annotations for tasks, actions, objects, rooms, hand-object interactions, task phases, and completion states.
View datasetMultimodal
Audio-Video Conversational Dataset
4,000 hours · 48kHz audio · 1080p+ video
4,000-hour global audio-video conversational dataset with synchronized speech, facial movement, mouth motion, gestures, turn-taking, emotion, laughter, interruptions, and multilingual speaker interaction.
View datasetPhysical AI
Multimodal Human Manipulation Dataset
6,000 hours · multi-sensor · mocap + RGB-D
6,000-hour sensor-rich human manipulation dataset with egocentric video, external cameras, RGB-D, motion capture, IMU signals, hand pose, object labels, task phases, and interaction annotations.
View datasetRobotics
Robot Teleoperation Dataset
2,000 episodes · 100-250 hours
2,000 robot manipulation episodes with multi-camera video, robot joint states, gripper states, action trajectories, object labels, task phases, and success/failure outcomes for robotics and Physical AI.
View datasetPhysical AI
Industrial First-Person Video Dataset
50,000 hours · 1080p+ · first person
50,000-hour first-person industrial and warehouse video dataset covering picking, packing, scanning, stocking, inventory, sorting, inspection, returns, assembly, cleaning, and workplace workflows.
View datasetWorld Models
Action-Conditioned Gameplay Dataset
10,000 hours · 500,000 episodes · 1B+ frames
10,000-hour rights-cleared gameplay dataset with synchronized video, keyboard/mouse/controller inputs, camera motion, player state, object state, events, goals, rewards, and outcomes for world models.
View dataset