Out of Browser courses

The eval you actually need

Stop reading leaderboards. Build the harness that measures your task, on your inputs.

What this course is

Every model decision you make — hosted or local, big or small, this quarter's or last — is a claim you cannot check without a way to measure your own task. This course builds that instrument, and nothing else.

It is free, all of it. It is the prerequisite for the other courses rather than a product.

What you'll build

  • A probe set drawn from your awkward inputs, not somebody's test split.
  • A scorer you trust, including for the cases where the right answer is "refuse".
  • A baseline number you can hold every later claim against.

Module 0 — Why leaderboards lie

Not because they're dishonest, but because they answer a question that isn't yours.

  1. The average task is not your task — A number that moved three points on a benchmark may move zero on your problem. (2 min)
  2. Three failures a benchmark will never show you — Refusals, near-misses and confident nonsense, none of which move an accuracy number. (2 min)

Module 1 — Tasks as data

Adding a probe should be appending to a list, never editing a runner.

  1. One record per probe — The shape that lets you add the hundredth test case as cheaply as the second. (2 min)
  2. Scoring without fooling yourself — Exact match is too strict, a model judge is too kind, and both are measurable. (2 min)

Related writing