RL Environments
RL Environments grounded in premium data
Train, test, and evaluate agents on expert-level tasks in our RL Environments. Built from licensed, non-public sources, so your models haven’t seen the answers.
Reward
0.62
Loss
0.41
Rollouts
1,024
Pass rate
41%
> env: coding.repair · grader: unit_pass
> rollout 0142 · tools: shell, editor, tests
> status: example run · illustrative values
How it works
Licensed data in trainable environments
01
Licensed at the source
Every task starts from non-public data, with rights cleared before anything is built.
02
Turned into tasks
Each environment ships with tasks, tools, a harness, and a scoring method.
03
Ready to train and evaluate
Run models, score with graders, and catch reward hacking and false passes.