Data

Complex human intelligence,
to fuel the frontier.

Code, multimodal, audio and speech, and trajectories, captured on our own stack from a panel of 981,000 humans, with versioned consent and full provenance.

Off the shelf

World class datasets available on demand.

Built on our own end-to-end pipelines. Active collections, delivered in weeks, in your format, to your cloud.

Custom collection

If you can describe it, we can capture it.

Talk us through your most complex data requirements. Where others have fallen short, we deliver.

tobymao/sqlglot @ d37583611 hidden checks8 attempts · 3 solved

Code

Real repositories, pinned to a commit.

Engineers work inside real repositories, pinned to a specific commit. Every task ships with a reference solution and is graded against hidden tests, rubrics and model judges.

Browse samples →

Active collections

  • TB4-SoftwareHarbor format software engineering tasks against SOTA models.2000 · ~3 wks
  • TB4-ScienceHarbor format science tasks against SOTA models.350 · ~6 wks
  • FrontierCode 2.0Code graded on merge quality against maintainer rubrics.1000 · ~6 wks
  • DeepSWEReal-repo issues resolved end to end, with tests and a reference.180 · ~3 wks
  • SWEQnA AtlasRepo-level questions pinned to a commit. Rubric and LLM-judge graded.
RecExpert cam
RecNovice cam
Heat map
Click path
Audio · 3 tracks
Audio
Screen capture
Transcription
Think-aloud
Timestamps
Handoffs
Consent
Click tracking
JSON schema
Provenance
Annotation
Translation

Multimodal

Real tasks, on screen and in the world.

Experts and everyday users complete real tasks on screen or in the real world while we record every step, thought, interaction and word. Sessions run from minutes to days, with prompts, assets and human references attached.

Browse samples →

Active collections

  • Paired TrajectoriesExpert user guiding novice user on computer use tasks.1500 · ~2 wks
  • Computer UseHobby-to-work computer use tasks across OS, file systems and code.300 · ~2 wks
  • Long-Horizon TasksRecorded screen sessions with matched video and instruction tracks.800 · ~3 wks
  • Egocentric TasksFirst-person video of household tasks, with a metadata schema.3500 · ~3 wks
Video · solo
Audio · pair
Video · pair
Audio · group
Audio · solo

Audio & speech

Unscripted speech, solo to multiparty.

Recorded solo, in pairs and in groups across up to 8 languages. Each session is segmented and transcribed, with a separate track per speaker.

Browse samples →

Active collections

  • SoloUnscripted solo sessions in specified languages.5000 · ~2 wks
  • Multiparty Conversational2+ speaker group conversations in a single specified language.1000 · ~2 wks
  • Multiparty, MultilingualMultiparty conversations with multiple languages and code-switching.750 · ~3 wks
  • Parallel MultilingualSame pair recorded in English and their native language.1000 · ~3 wks
Professional · brief
Assistant · draft
MDoverride.md· engineering−▶···

          

          


          
Quality 

          
Effort 

          
Accepted · with edits

          

        

        

          
Trajectories

          
Professionals working with an AI assistant.

          
Real tasks, captured turn by turn through to the final output, with structured ratings on quality and effort.

          
Browse samples →

          
Active collections

          
    
            
  • Multi-turn PromptsProfessionals completing real world work, with all turns and outputs.500 · ~1 wk
    • 
            
  • Documents - LegalLegal practitioners drafting commercial documents one to one with an AI assistant.2000 · ~1 wk
    • 
            
  • EngineeringEngineers share AI prompt override .md files and explain the reasoning.1500 · ~3 wks
    • 
            
  • Document CreationProfessionals drafting documents. Full session, tool calls and output.On request
    • 
            
  • Document CollectionVerified, owned documents and context from consenting professionals.On request
    • 
          

        

      


    

  


  

      
      

        

          

            
Custom

            
If you can describe it, we can capture it.

            
Our own end-to-end stack captures anything in high fidelity: video, audio, screen and text, millisecond aligned, with reasoning, handoffs and provenance.

            
Deliveries typically include

            
    
              
  • ModalitiesVideo, audio, screen and text
    • 
              
  • AlignmentMillisecond timestamps across modalities
    • 
              
  • ReasoningThink-aloud and post-capture annotation
    • 
              
  • Unit of dataSession or task, single or multi-stage
    • 
              
  • HandoffsTiming and context between participants
    • 
              
  • ProvenanceParticipant, brief, date, consent and demographics
    • 
              
  • AssetsSynced video and audio, transcripts, JSON schema
    • 
            

            
Request data →

          

          

            

              

                
                

                
Spec

                
Match

                
Consent

                
Capture

                
Process

                
Validate

              

            

          

        

      


  


  

    

      

        
Tell us what you need, we'll make it happen.

        
Start with around 10 sample tasks, then a pilot of 50 to 100, then production in batches to your cloud, in your format.

        
Browse samples →Request data →