FLAN (instruction-tuned baseline)
Training language models to follow instructions with human feedback — inherited
A compilation of NLP tasks with natural-language instructions (Wei et al., 2021) that the paper fine-tunes 175B GPT-3 on to produce a cross-task-instruction-tuning baseline for comparison against InstructGPT.
One way to teach a model to follow instructions is to compile hundreds of NLP tasks, rewrite each with a natural-language instruction, and fine-tune on the lot. FLAN is that approach, and InstructGPT trains a version of GPT-3 on it precisely to have a rival to beat.
InstructGPT builds this baseline deliberately, and what it enabled, plus the gap it exposed, matters more than the recipe itself. T0, built the same way, is worth reading alongside it.
How it fits
FLAN (Wei et al., 2021) is a compilation of NLP tasks reformatted with natural-language instructions, inherited into InstructGPT as one of two prior cross-task instruction-tuning baselines — the other being T0 — inside the Instruction following as an alternative to next-token prediction theme's argument for why direct RLHF training beats fixed-compilation instruction tuning. InstructGPT fine-tunes 175B GPT-3 on FLAN to produce a comparison baseline, rather than treating FLAN's own released checkpoints as the comparison point (instructgpt, §"1 Introduction", p. 4; §"C.5 FLAN and T0 models", p. 42).
What it enabled: a controlled baseline, and a diagnosed gap
Before it could be compared fairly, FLAN first had to be put on equal footing with its sibling baseline: T0 is scaled to match FLAN. T0's T0++ variant contains roughly 96 million datapoints against FLAN's roughly 1.2 million, so InstructGPT's authors subsampled T0 down to about 1 million examples before fine-tuning, and both resulting models were selected by the same checkpoint-picking rule as the SFT model — highest reward-model score on the validation set rather than lowest validation loss — tying all the comparison baselines to one consistent selection criterion (instructgpt, §"C.5 FLAN and T0 models", p. 42). Once comparable, FLAN's role becomes diagnostic: it sets the comparison floor for instruction following. InstructGPT outputs are preferred over FLAN's 78% of the time head-to-head, and the paper offers a specific structural explanation rather than leaving the gap unexplained — FLAN's fixed, pre-formatted task compilation emphasizes tasks that are easy to score automatically, classification and question answering, which together make up only about 18% of real API traffic, while open-ended generation and brainstorming, poorly represented in FLAN, account for roughly 57% (instructgpt, §"4.1 Results on the API distribution", p. 13). FLAN's ceiling is therefore attributed to what it was built from, not to any deficiency in the fine-tuning procedure applied to it.