The data frontier models learn from.
ARIMLABS is a frontier training-data research company. We build the RL environments and post-training data frontier models train on, and we study what that data does to a model: how capable it becomes, and how safe.
why we exist
Frontier models are limited less by compute than by what they are trained on. The next capability gains come from unique and valuable post-training data.
We also think the field knows too little about what its data does. A dataset goes in, a model comes out, and the effect in between is a black box. That limits how predictable capabilities and safety are.
who we are
A small team with an unreasonable track record.
- 01Top-ranked CTF players.
- 02Winners and laureates of international olympiads in informatics and mathematics.
- 03Security researchers with CVEs against the software the world runs on.
All of us believe in an abundance of intelligence and what it does to the economy. All of us are here to make it arrive faster, and safer.
Mykyta Mudryi
ceo · x ↗AI security and safety researcher and a top CTF competitor since age fifteen. Contributed to Google DeepMind's open-source security work, which fixed 13,000 vulnerabilities.
- Google DeepMind
- CTF since fifteen
- AI security & safety
Markiian Chaklosh
coo · x ↗Previously at Akamai, with six years of industry experience.
- Akamai
- Operations
- Six years in industry
what we do
We curate environments, built from human data and in combination with synthetic data, and we ask the most pressing questions about that data through the research around it.
- 01
Human and synthetic data generation
How far can synthetic generation stretch expert-authored data without flattening it?
We study pipelines that scale coverage while holding difficulty, high-fidelity, and diversity at expert level.
- 02
Reward design and verification
What does a verifier need to measure the skill rather than its proxies?
We build reward signals that credit partial progress, hold up under optimization pressure, and close off reward hacking.
- 03
Data composition and learning efficiency
Which mix of isolated skills and end-to-end scenarios, task difficulty, and horizon length produces the most capability per gradient step?
We treat curriculum as a measurable design variable, not a guess.
- 04
Agent generalization and evaluation robustness
When does an agent solve the task, and when does it solve the benchmark?
We examine how environment fidelity, evaluation awareness, and deployment scaffolds shape shortcut-taking and transfer to the real thing.
- 05
Post-training dynamics and safety
What does a dataset actually do to a model?
We track held-out gains, skill regressions, and safety behavior across compute budgets, so the effect of training data is measured rather than assumed.
work with us
We take on the hardest problems on the way to an abundance of intelligence. If you need high-quality data for post-training, let us know.