Measuring what you mean

You want to study "understanding", "helpfulness" or "learning". But you can only measure something concrete: a test score, a rating, a benchmark number. The gap between the two is where many experiments quietly fail.

Two qualities of a measure

• Validity: does it measure what you mean? A typing-speed test is not a valid measure of writing quality.
• Reliability: does it give consistent results? A rating that changes every time the same person rates the same answer is unreliable.

A measure can be reliable but not valid: a bathroom scale that always reads 2 kg too high.

When the measure becomes the target

Economist Charles Goodhart observed that when a measure becomes a target, it stops being a good measure. In ML, models tuned hard on one benchmark can learn shortcuts that raise the score without the ability the benchmark was meant to test. For example, studies of natural-language inference datasets found that models could often guess the answer from the hypothesis sentence alone, without reading the premise.

Good habits

• Write an operational definition: exactly how the outcome is measured.
• Use more than one measure for big ideas.
• Check reliability: have two people rate the same items and compare.
• Look at examples, not just the number.

RememberDecide exactly how you'll measure the outcome, and check that the number really means what you want it to mean.