DatoricProducts
Licensed voice and video,
cleared at the source.
Licensed, ethically sourced multimodal data for frontier AI development. Captured directly from consented native speakers and domain experts, with fair compensation and clean rights end to end.
Voice Data
Licensed voice data recorded by consented native speakers and domain experts, with fair compensation and clean rights.
We work directly with thousands of native speakers and credentialed practitioners under explicit consent and fair compensation agreements. From low-resource languages to highly technical reasoning, every recording is sourced at the origin, rights-cleared, transcribed, and validated by the people who actually speak it.
Capabilities
- →Multilingual collection across 50+ languages
- →Expert reasoning traces, medical, legal, technical
- →Conversational, narrated, and instructional formats
- →Native-speaker evaluation benchmarks
Video Data
Licensed multimodal video at the resolution of expert work, for vision-language and agentic models.
Our video data captures how experts actually do their work, click by click, decision by decision, with the audio narration and text artifacts that ground each action. Every contributor is consented, fairly compensated, and credited under a clean license. The result is paired multimodal data that teaches models to reason, not just describe.
Capabilities
- →Workflow demonstrations and tool-use traces
- →Multi-modal alignment across vision, audio, and text
- →Truth and hallucination evaluation suites
- →Domain-specific corpora at scale
Computer & Browser Use Data
Consented expert workflows captured click by click, across browser and desktop, for agents that use real tools.
Public computer-use benchmarks lean heavily on static HTML, single operating systems, and tasks without recovery paths, which is why frontier agents still fail on real software. We capture consented expert workflows in motion: the DOM as it shifts, the accessibility tree as it rebuilds, the click that misses and the correction that follows. Every trace is licensed, rights-cleared, and step-annotated, so agents can learn from the web and the tools experts actually use.
Capabilities
- →Paired screenshots and action traces — click, scroll, type, drag
- →DOM snapshots, accessibility trees, HAR logs, MHTML archives
- →Cross-platform coverage across browser, macOS, Windows, Linux
- →Multi-step workflows with error, recovery, and undo paths
Datasets
Available datasets
Licensed, consent-sourced corpora ready for evaluation and licensing. Request access to review samples and pricing.
Alignment
RLHF-Voice
On-policy voice preference pairs for model alignment. Human evaluators rank model outputs with spoken rationales.
Expert Data
ExpertTrace
Credentialed practitioner reasoning traces across medical, legal, and financial workflows. Step-by-step decision chains with verbal narration, capturing the judgment current models cannot learn from web-scraped text alone.
Voice
GlobalASR-50
Multilingual ASR corpus spanning 50+ languages with accent, dialect, and recording-environment labels. Goes beyond single-language collections with cross-lingual speaker metadata and phonetic annotations.
Agentic
AgentTrace-Web
Browser and desktop agent interaction traces with DOM snapshots, accessibility trees, and full action sequences. Expert workflows capturing tool use, error recovery, and multi-step task completion across real applications.
Voice
LowResource-Voice
Native-speaker speech for underserved languages and dialects where no public corpus exists. Covers languages across South Asia, Sub-Saharan Africa, and Southeast Asia that current foundation models cannot transcribe.
Video
ExpertWorkflow-Video
Expert-narrated workflow video with synchronized audio commentary and step-level annotations. Domain practitioners demonstrating real procedures in software, clinical, and technical settings, not scripted reenactments.
Expert Data
MedExpert-Voice
Healthcare voice data from verified practitioners: clinical reasoning narrations, diagnostic walkthroughs, and patient-facing explanations. Voice-native medical data, not text records converted to speech.
Expert Data
LegalExpert-Voice
Legal domain expert voice recordings covering contract analysis, regulatory interpretation, and case reasoning. Sourced from practicing attorneys with jurisdiction-specific terminology and argumentation patterns.
Multimodal
MultiModal-Align
Synchronized vision-language-audio alignment pairs where all modalities are captured together, not stitched after the fact. Expert narration paired with video and text artifacts for cross-modal grounding.
Talk to us
Tell us what your model is missing.
We work with frontier labs and applied teams on custom datasets, evaluation suites, and ongoing data partnerships.