Types of data
Capture, preference labels, and expert evaluation Harbor actually delivers.
Harbor delivers three kinds of human data. Programmes are scoped with you. There is no self-serve catalog.
Multimodal capture
First-person video, spatial audio, and sensor streams from real commercial environments. Contributors record work they already know how to do. Consent and identity attach at capture through Harbor Passport.
Typical settings: warehouse, trades, logistics, clinical workflow (operations only — not patient records), and residential tasks.
Preference and evaluation data
Experts score or rank model outputs against a published rubric. This is the human feedback used for RLHF, DPO, and reward models: pairwise ranking, Likert scores, and written ideal completions where the programme needs them.
This is not generic labeling. The task is judgment: is this grasp plan feasible, is this summary safe, does this reply break policy.
Expert review of capture
Domain specialists check capture quality: completeness, task accuracy, framing, audio, safety, and policy. Gold items and calibration keep reviewers aligned.
What Harbor does not sell as a default
Harbor is not a commodity image mill. Programmes that mix real and synthetic rare-class footage are scoped as managed work, not a public SKU.
Talk to Harbor at book a conversation if you need a mix of capture and eval.
Need a programme scoped? Talk to Harbor.