Datasets
Real-world training data across voice, video, 3D, and paired text. Collected with consent, licensed for commercial use.
Video
8Real-world video: interviews, egocentric tasks, and robots at work.
Robotic Arm: Pick-and-Place & Manipulation
Robotic-arm pick-and-place episodes pairing 1280x720 video at 30 fps with synchronized robot telemetry, camera information, and manifests, so policies can be evaluated against ground-truth sensor state rather than video alone.
Cultural Interviews: Public Conversations & Perspectives
Public interview videos with unscripted cultural conversations, shot in natural settings at 1080p. Metadata covers interview setting, conversational style, and framing, so coverage can be reviewed before scoping more.
Egocentric Video: Cleaning, Laundry & Car Wash
First-person task videos covering appliance cleaning, car wash and repair, interior work, and pet sitting, filmed from the worker's own viewpoint at 1080p and 30 fps with hands, tools, and task flow in frame.
Humanoid Robot Training: Egocentric Manipulation
Egocentric video of real-world manipulation for robotics and embodied AI, captured at 1080p and 60 fps through a super-wide lens, close to what a humanoid platform's head camera sees during household work, assembly, and collaboration tasks.
First-Person Task Video: Household & Maintenance Activities
Curated first-person videos spanning kitchen work, gardening, light maintenance, cleaning, and fine-motor manipulation, with clear hand-object interaction in every clip and per-clip metadata for category, task type, and capture properties.
Warehouse: Kitting, Pick-and-Place & Packaging
Human workers performing kitting, pick-and-place, and packaging tasks in a simulated warehouse, captured from third-person and top-down views at 1280x720 with synchronized audio for logistics and robotics evaluation.
Behavioral Dataset: Multimodal Video, Depth & Actions
Human-teleoperated demonstrations of everyday household tasks that synchronize RGB video with quantized depth and instance segmentation from head and wrist cameras, shipped in LeRobot format so they drop into existing robot-learning pipelines.
Computer Use: Agent Operating a Desktop Interface
Agent-view recordings of desktop interaction captured at 60 fps, with cursor movement and GUI actions visible frame by frame and subtitle tracks marking on-screen task execution for computer-use agents.
Audio
12Speech and dialect audio recorded by native speakers, each clip paired with a transcript.
Bengali Speech: Single & Multi-Speaker
Bengali speech spanning single-speaker narration and two-speaker conversations, so models see natural turn-taking as well as clean prompts. Audio ships as mono 44.1 kHz WAV, and every clip carries a ground-truth transcript with speaker structure and format metadata.
Hindi Speech: Contributor Recordings
Hindi speech recorded through a moderated contributor app, keeping capture quality and formatting consistent across speakers. Audio ships as mono 48 kHz WAV, and each clip pairs the recording with a validated transcript plus duration, codec, and speaker metadata.
Spanish Speech: Contributor Recordings
Spanish speech recorded through a moderated contributor app, keeping capture quality and formatting consistent across speakers. Each clip pairs the recording with a validated transcript plus duration, codec, and speaker metadata.
Korean Speech: Contributor Recordings
Korean speech recorded through a moderated contributor app, keeping capture quality and formatting consistent across speakers. Audio ships as mono 48 kHz WAV, and each clip pairs the recording with a validated transcript plus duration, codec, and speaker metadata.
Urdu Speech: Contributor Recordings
Urdu speech recorded through a moderated contributor app, keeping capture quality and formatting consistent across speakers. Audio is mono at 48 kHz, and each clip pairs the recording with a validated transcript plus duration, codec, and speaker metadata.
French Speech: Contributor Recordings
French speech recorded through a moderated contributor app, keeping capture quality and formatting consistent across speakers. Audio ships as 48 kHz WAV in mono and stereo, and each clip pairs the recording with a validated transcript plus duration, codec, and speaker metadata.
Bengali Speech: Read & Conversational
Bengali speech captured across read prompts, narration, and free conversation, so the same voices appear in both controlled and natural speech. Audio ships as WAV at 44.1 and 48 kHz with a ground-truth transcript on every clip, ready for ASR, voice interfaces, and alignment work.
Tamil Speech: Read & Conversational
Tamil speech covering read and conversational sessions side by side, which matters for a language where the spoken form drifts far from written text. Audio ships as 48 kHz WAV with ground-truth transcripts on every clip, built for ASR and low-resource language research.
Telugu Speech: Read & Conversational
Telugu speech spanning scripted prompts, narration, and conversational samples, so models can be checked against both careful and casual speech. Audio ships as 48 kHz WAV, and each recording is paired with a transcript for ASR evaluation and voice assistant workflows.
Chinese (Mandarin) Speech Samples
Mandarin speech from native speakers, with each clip pairing mono 48 kHz audio with a ground-truth transcript, language labels, and speaker context for speech recognition and voice interface evaluation.
Chinese (Cantonese) Speech Samples
Cantonese speech from native speakers, with each clip pairing mono 16 kHz audio with a ground-truth transcript, language labels, and speaker context, a match for common telephony and ASR pipelines.
Chinese (Henan) Speech Samples
Henan Chinese speech from native speakers, a regional variety rarely available as labeled data. Each clip pairs mono 16 kHz audio with a ground-truth transcript, language labels, and speaker context for regional speech analysis.
3D
13D scans and motion capture recorded on studio rigs.
Pairs
2Question-answer and code pairs for models that reason step by step.
Q&A Pairs: Physics, Chemistry, Math & Biology
Verified Q&A pairs across Physics, Chemistry, Math, and Biology, aligned with JEE, NEET, and CET-style exams. Every pair is tagged with subject, difficulty, curriculum, and question type, and image-based questions keep their diagrams.
Code I/O Pairs: Prompt-Driven Few-Shot Workflows
Developer workflow pairs where each row aligns a prompt with the code it produced, the tests it ran, the error traces it hit, and the corrected output, drawn from real workflows rather than synthetic generation.