Training & evaluation data
Koyalab produces verified datasets for coding models, AI agents and embodied systems — built by specialists and checked by execution wherever the result can be run.
/ What we build
Code that has to run, not just read well.
Professional engineers write and repair real software tasks — from single functions to repository-scale changes. Every solution ships with the tests that prove it, and every review explains why one patch beats another.
✓ executed in sandbox · tests pass · reviewer sign-off
Trajectories for models that take actions.
We record and grade complete multi-step episodes: an agent reading a goal, calling tools, browsing, operating software and recovering from mistakes. We also build the environments and reward rubrics those agents are trained against.
✓ replayable episode · per-step labels · outcome verified
Data for robots and the physical world.
Teleoperated demonstrations, egocentric video and dense action annotation for manipulation and navigation. Collected with trained operators under controlled protocols, segmented into reusable skills, and aligned across sensors.
✓ synced sensors · protocol logged · QA sampled
/ How a project runs
We start from your model's failure cases and agree on a written spec: task mix, difficulty, format and acceptance criteria.
We set up the environments, rubrics and tooling the work needs — sandboxes for code, harnesses for agents, rigs for robots.
Vetted specialists produce the data in small, reviewed batches so the spec can tighten as early results come in.
Automatic checks run first, then independent human review. Only units that pass both are delivered, with their full audit trail.
/ Sample unit
Every step is recorded, replayable and labelled. Mistakes are kept and marked, because recovery is part of what a model needs to learn.
Where a result can be executed, replayed or measured, we check it that way before a person ever signs it off.
Engineers, researchers and trained operators who understand the domain — each accountable for the work they submit.
Datasets are built to order for your team, delivered in your format, and never resold.