All tiny agents
Evals
Reward Grader
Naive vs composite RL reward with a reward hacking detector.
What it does
Grades six scripted attempts at one bug fix with a naive reward and a composite reward zeroed by a hack detector, then computes GRPO-style group-relative advantages. No model.
The guardrail
A hack detector that zeroes gamed rewards.
Run it
git clone https://github.com/bobbyhalljr/tiny-reward-grader.git && cd tiny-reward-gradernpm installnpx tsx reward.ts- Repo
- tiny-reward-grader
- Language
- TypeScript
- Stars
- 0
- Price
- Free