FLAN (instruction-tuned baseline) — sets the comparison floor for → Instruction following (as a fine-tuning objective)

InstructGPT outputs are preferred over FLAN's 78% of the time head-to-head (instructgpt, §"4.1 Results on the API distribution", p. 13), and the paper offers a specific diagnosis for the gap rather than leaving it as a bare number: public NLP compilations like FLAN are built around tasks that are easy to score automatically — classification and question answering — which together make up only about 18% of what real API customers actually ask for, while open-ended generation and brainstorming account for roughly 57% (instructgpt, §"4.1 Results on the API distribution", p. 13). FLAN's fixed, pre-formatted task compilation therefore systematically underrepresents the organic, open-ended prompt distribution instruction following is trained and evaluated on, which is presented as the structural reason a fixed-compilation approach cannot match a broad, real-usage-derived training signal.