Ablation
An experiment that removes or changes one component of a model or training setup to measure how much it actually contributes.
Borrowed from biology, an ablation isolates one design decision — a dataset, an architecture tweak, a training trick — by training with and without it and comparing performance on evaluation benchmarks. The difference is attributed to that component.
Good ablations change one thing at a time and are usually run at smaller scale to keep costs manageable, which introduces a risk: a component that helps at small scale may not help at large scale, or vice versa.