Quick start

Every experiment has a --tiny mode: a small model trained to show the behaviour is built on your CPU, so the whole pipeline runs without a download. Its numbers mean nothing; it checks the path.

uv run --all-extras python experiments/refusal-direction/run.py --tiny
uv run --all-extras python experiments/sycophancy-pushback/run.py --tiny
just serve   # http://127.0.0.1:8000

Runs lists both; Examples walks through what each figure says. A run opens on its overview, with figures a tab away; an eval run's Samples tab opens every transcript; Compare puts two eval runs side by side with a paired interval.

On a real modelCopy Anchor Link

Drop --tiny. The default is Qwen/Qwen2.5-0.5B-Instruct, which runs on a laptop; a GPU makes 1.5B and up practical.

uv run --all-extras python experiments/sycophancy-pushback/run.py --model Qwen/Qwen2.5-1.5B-Instruct

Your own questionCopy Anchor Link

just new-experiment my-question   # experiments/my-question/{README.md,run.py}

The README states the question, what would answer it and the result. run.py writes to MLflow and Inspect logs, never to stdout alone, so the app shows it.