Here are the hands-on walkthroughs for running benchmarks, collecting rollouts, and reading the outputs. They assume familiarity with the basic concepts in Evaluation and the workflow in Evaluation.
Run the EvalPlus coding benchmark and inspect rollout and aggregate metric outputs.
Run BabyVision with OpenCode in a sandbox session that has a Python image toolkit, graded by a GPT judge.
Generate one validation task per probe, then inspect a multi-turn conversation and score.
Recompute rewards from existing rollouts after changing a verifier parameter, without re-running model inference.
Run an agentic tool-use benchmark end to end with the Claude Code agent harness — config, rollouts, and BLADE analysis.
notebookBrowse the built-in benchmark and training environments.
Understand the aggregate metrics written after rollout collection.