Verifiable environments built from real workflows, agent trajectories captured at our own gateway, and expert judgment on what good looks like. Pretraining made models smart. This is the data that makes them useful.
Every record starts with a real task, a real environment and a person or an agent doing the work. We turn that into data a model can be trained and scored on.
Real workflows rebuilt as interactive environments: repository tasks with tests, e-commerce operations, customer support, logistics dispatch. Where the business has a ground-truth outcome the environment scores it; where it does not, expert rubrics do. Delivered with fixtures, reward functions and held-out splits.
Complete multi-turn runs of coding agents captured at Realset's own LLM gateway: system prompt, tool schemas, every tool call and result, reasoning where the model emitted it, final answer. Sliced by model version, task type and date; de-identified at source with per-file integrity hashes.
Domain experts write rubrics, rank outputs and diagnose failures against real workflows. Preference pairs and rubric scores feed reward models; failure diagnoses become targeted fix-data. Human review beyond LLM-as-judge.
Egocentric and teleoperation capture in homes, kitchens, warehouses and assembly lines, with 3D hand pose, camera calibration and action-level labels. Delivered in LeRobot or your schema, with real-world evaluation trials for the policies you train on it.
One pipeline for environments, trajectories and expert judgment, with a verifier at every step.
Task family, environment, what counts as done: tests, business outcomes or an expert rubric.
Agents run in the environment through our gateway; experts and skilled workers produce demonstrations and judgments. Our AI recruiter screens them for the exact task.
Tests run, rubrics are scored, an independent reviewer checks every record. Failures are named and counted, not averaged away.
JSONL, LeRobot or your schema, de-identified, hashed, with a live pipeline dashboard. The next batch targets what your model got wrong.
Every record ships with the evidence that it was checked. Buyers get the numbers, not a promise.
Coding tasks carry fail-to-pass and pass-to-pass tests. Business tasks carry the real outcome. Judgment tasks carry a written rubric scored item by item. Nothing enters a batch on a reviewer's feeling.
The reviewer is never the producer. In the Workspace they see the answer, the evidence and the screen recording of the work, and record a decision that stays attached to the record.
Tool calls paired with results, schema conformance, duplicate-turn ratio, truncated or interrupted runs flagged, tool names normalized against the original. Sessions that fail are excluded and counted.
A manifest lists every object with its size and hash. You verify the delivery yourself; a 793-object embodied pack is checked file by file before it leaves our storage.
Emails, secrets, IPs, paths, usernames, organisation and person names are replaced with consistent placeholders before data leaves our storage. Each record reports how many replacements were made, so a reviewer can see what was touched.
Each delivery includes metadata.json: per-session model, turns, tool calls, usage coverage, what was captured and what was not. Failures are named and counted, never averaged away.
Where domain experts answer real tasks, record how they got there, pass independent review, and get paid. Every approved record exports as JSONL with provenance you can verify.
Experts contribute an answer, supporting evidence and, when the task calls for it, a screen recording of the work, stored with a SHA-256 digest.
A separate reviewer scores correctness, evidence and completeness against the rubric. Only approved records leave the workspace.
English, Simplified Chinese, Japanese, Korean, Spanish, Arabic. Payouts in USDT, USDC, BTC, IBAN, PayPal, Wise or Alipay, tracked on a ledger.
We publish open benchmarks on real-world tasks so the field can measure what matters: whether a policy works outside the lab.
Open-source VLA policies evaluated on household tasks captured in real kitchens and laundry rooms, not simulation.
| Tasks | Folding, loading, sorting, wiping |
| Policies under test | π0, OpenVLA, GR00T, Octo |
| Metric | Success rate, 3 trials per task |
Bimanual assembly and packaging tasks from real production lines, scored by the line workers who trained on them.
Join as a partner lab →Computer-use agents on real e-commerce seller operations: listing edits, ad bids, returns and review handling.
Join as a partner lab →Three questions. We reply within one business day with a sample set and a quote.