CH 04 · Glossary: Benchmarks · 06/13
Read as one pageTask and task suite
A task is one specific kind of problem you ask a model to solve, for example translating a sentence or answering a multiple-choice science question. A task suite is a bundle of many such tasks gathered into one benchmark. The reason large models are tested with suites rather than single tasks is that a generalist model can be strong in one area and surprisingly weak in another, and only a wide spread of tasks reveals the full shape of its abilities. BIG-Bench is a task suite of 204 tasks, deliberately spanning many different domains.
Finished this idea?
Your learning progress stays in this browser.