Definition
Plain language
Drawing the smooth line that best matches a set of measured points.
As stated in the literature
Estimating parameters of a functional form from data; flexible forms with free parameters (e.g. an unconstrained ceiling) can fit many monotone datasets, so goodness of fit is weak evidence for the underlying mechanism.
Also called: fit the curve, fitting
Why it matters: A line that hugs the data points can still be the wrong explanation of what is happening, so a good fit alone should not be read as proof that you understand the mechanism or can extrapolate beyond the measured range.
For example, if you measure a model's error rate at five different training sizes and draw a smooth declining line through those five dots, you are curve fitting.
Heard on the show
“The loss stays essentially flat — the model is fitting the documents just as well as before.”Episode 043 — When 'This Is False' Doesn't Stick: Why Models Learn the Lie Anyway