AI training data and evaluation
Training data
built by
domain experts.
Oxyriz is an AI data services and domain expert provider. From raw information to real intelligence, we source, clean, and build the expert data behind more capable AI.
Data sourcing · Training datasets · Custom evaluations
Authorized data, sourced to your brief.
Relevant business data, sourced through authorized partnerships.
- Coding
- Finance
- STEM
- Robotics
- Human feedback
01 / What we do
Data sourcing, training data
and expert review.
Specialist knowledge. Carefully prepared data. A clear path from your challenge to delivery.
Source & prepare
Authorized business data, cleaned and structured around your requirements.
Train & evaluate
Bespoke training datasets and evaluation benchmarks, authored by working specialists.
Bring in expertise
Domain experts who bring real-world judgment to your most demanding AI challenges.
02 / From theory to practice
Robotics tasks that
stump Fable 5.1.
Our MuJoCo robotics simulations. Complex challenges, observable outcomes.
Adaptive navigation
Snow-course navigation
A vehicle follows a winding course through snow, low-traction ice, and rock obstacles. The recorded run reports 24.74 m of progress and zero collisions.
- Oracle
- 1.00 Pass
- Reference
- 0.50 Pass
- Fable 5.1
- < 0.10 avg
- No-op
- 0.00 Floor
Contact-rich control
Underwater ROV commissioning
An underwater vehicle visits five ports, makes probe contact, completes a coded handshake, then retracts the probe and holds position.
- Oracle
- 1.00 Pass
- Reference
- 0.50 Pass
- Fable 5.1
- < 0.20 avg
- No-op
- 0.00 Floor
How a task earns its place
Fair enough to solve.
Hard enough to matter.
Every run is scored from 0 to 1 against fixed anchors. A task ships only when the oracle and reference pass, doing nothing scores zero, and frontier models stay under 0.40. Fable 5.1 averages below 0.10 on snow-course navigation and below 0.20 on ROV commissioning.
- 01
Oracle sets the ceiling
A controller with privileged simulator state scores 1.0, proving the task is physically solvable.
- 02
Reference proves it’s fair
An expert controller seeing only what the model sees scores 0.5. No hidden information is required.
- 03
No-op sets the floor
Doing nothing, or a naive policy, maps to 0. A score has to be earned through control.
- 04
Hidden scenarios
Grading uses held-out scenarios sampled from ranges disclosed in the brief, so memorising the test doesn’t help.
- 05
Frozen before evaluation
Scenarios, scorer and anchors are locked before any model runs and never re-tuned against one.
- 06
Sandboxed and verified
Submitted policies run in an isolated process, with physics-integrity checks that block shortcuts.
The videos show demonstration rollouts from our controllers, not recordings of Fable 5.1’s attempts. Scores are normalised per task: no-op 0, reference 0.5, oracle 1.0.
03 / How it runs
A project, end to end.
Tight feedback loops with your research team, from first brief to continuous delivery.
- 01
Brief
We meet your research leads to scope the failure modes you want closed, the schema you need, and the success criteria you’ll grade against.
- 02
Source & calibrate
Domain experts are staffed from our network, calibrated on a gold set, and onboarded into the project-specific rubric.
- 03
Deliver & iterate
Datasets ship in your preferred schema with rubric scores and reviewer trails. Continuous review keeps them current as your models evolve.
How do I start a data project?
Tell us your domain, goals, expected deliverables, and timeline through our contact form. Our team will review the requirements with you.
Can you source and clean business data?
Yes. We work with authorized business data through our partner network, with sourcing, processing, and cleaning tailored to the project.
How can I join the expert team?
Visit Careers, open the full role requirements, then select Apply. You can submit your details and resume through the application form.