rewardkit
Give Harbor tasks checks that produce scores.
What it does
Creates Harbor task verifiers that score results using programmatic checks or model-based judges.
When to use it
Use when creating or editing a task's tests, adding grading criteria, configuring judges, or designing reward-scoring verifiers.
How to use it
Give it a Harbor task to equip with grading checks; it writes verifier criteria and configuration into the project.
What you provide
- An existing project
Uses
What you need · 4
The verifier entry point specifies Python >=3.12,<3.14.
The guide targets Reward Kit; its verifier entry point selects harbor-rewardkit==0.2.* through uvx.
For judge criteria needing this API key, the example forwards ANTHROPIC_API_KEY from the environment through task.toml.
The optional JEV judge backend requires TYPESAFE_API_KEY; the files do not specify how the key is retrieved.
Access · 0
No additional actions listed.