Reinforcement learning environments
Long-horizon coding environments for frontier labs.
We build the tasks and graders that teach models real machine learning engineering: work that takes an expert hours in a real codebase, scored by a grader that only rewards doing the work.
What we build
RL environments
Tasks and graders for machine learning engineering: debugging a training pipeline, tracking down data leakage, fixing a broken evaluation, speeding up a slow model.
Evaluations
Every task is run against the strongest models available. We report how often each one solves it, so a lab can see when a new model improves.
Expert engineers
Each task is written and reviewed by working ML engineers, recruited and paid through Kirigi, our expert platform.
How we build an environment
-
01
Written by ML engineers
Each task comes from a working engineer and is based on a problem they have run into on the job.
-
02
Graders we try to break
Before a task ships, we try to beat its grader without doing the work. If we find a shortcut, we fix the grader.
-
03
Checked against the best models
We keep the tasks today's strongest models solve only some of the time, which leaves room to measure progress.
-
04
Long, realistic tasks
A task takes an expert hours in a real codebase. Current models still find that kind of sustained work hard.
Training or evaluating a model?
Tell us what you want your model to get better at. We will send sample tasks and how we would build an environment for it.
Are you an ML engineer?
Write and review these tasks, remotely and paid per task, through Kirigi, our expert platform.