profile pic
⌘ '
raccourcis clavier

Source

The paper argues that zero-shot prompting on a instruction-tuned small language models outperform LLMs systems.

  • Instruction-tuning actually improve zero-shot learning performance.
  • Mostly tested on FLAN, but show results throughout with GPT-3 and on few reading comprehension dataset.

Honorable mentions include prompt tuning or few-shots prompting