Twice the data, less power: my stability rule got blinder the harder I looked
Originally published at erikhill.dev . The numbers below are checked against the repository they come from. Twice the data, less power: my stability rule got blinder the harder I looked This is a finding about a measurement rule, not about a model. The models did not change. What happened I was comparing two models on a frozen 159 task suite and I had a problem most eval code does not admit to…
After comparing two models on a frozen 159 task suite, the same model sometimes disagreed with itself when asked the same question multiple times. To address this issue, a rule called "strict" was implemented, which discarded tasks where the config did not agree with itself. In the first three repetitions, strict gave a result of Haiku ahead 7 to 1, with 8 informative tasks and a p-value of 0.070, which was not significant.
However, upon further investigation, 13 tasks were discarded by strict, leading to the creation of a second rule called "rate." This rule tolerated a minority disagreement and resulted in Haiku ahead 13 to 2, 15 informative tasks, and a p-value of 0.0074, which was significant. The conservative rule was used in the beginning, leading to a null result, and then the rate rule was built to return the desired result. This is a common practice among those who "p hack" themselves.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
