Course in preparation · Coming soon
RL Environments, Graders and Verifiers
Build what a model is trained against. Design a task environment, write a grader the model cannot game, measure how often it is gamed anyway, and train a small model against it on Colab.
Who it is for. Engineers and researchers aiming at post-training and RL teams, where postings ask for environments, verifiers and reward design.
The planned syllabus
Planned lessons will help you understand each method, apply it to an AI problem, and judge whether the result holds up.
- 01
What an environment is
State, actions, episodes and the reward, for a coding task and a tool-use task.
- 02
Verifiers and graders
Exact checks, test suites and model judges, and the error rate of each.
- 03
Reward hacking
Seed shortcuts on purpose and measure how often a model finds them.
- 04
Training against the environment
A small policy-gradient run on a tiny model, with the curve and the failures.
- 05
Environments at scale
Sandboxes, parallel rollouts and the cost of each episode.
