Finetuned Language Models Are Zero-Shot Learners

The problem: doing a task with no examples

Zero-shot learning means a model has to carry out a task it has never been trained on, with no worked examples in the prompt. A pretrained language model has mostly learned to continue text, which is not the same thing as doing what a user's prompt asks for. The paper sets out a simple method for improving the zero-shot abilities of language models, and tests whether such an ability can be taught rather than only acquired through scale ^[1].

The authors are Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai and Quoc V. Le ^[1].

The method: instruction tuning

The central idea is what the paper calls instruction tuning: finetuning a language model on a collection of tasks that are described to it via instructions. Rather than training on raw text, the model is trained on many tasks in which a natural-language instruction states what should be done with the input that follows. The paper reports that finetuning in this way substantially improves zero-shot performance on tasks the model has not seen ^[1].

In the concrete experiment, the authors take a 137B parameter pretrained language model and instruction-tune it on over 60 NLP tasks, each verbalised through natural language instruction templates. The resulting instruction-tuned model is called FLAN ^[1].

Results

FLAN substantially improves on its unmodified counterpart. It also surpasses zero-shot 175B GPT-3 on 20 of the 25 tasks the authors evaluated, which means an instruction-tuned model smaller than that comparison model comes out ahead on most of the benchmark set ^[1].

The gains are not limited to matching a zero-shot baseline. FLAN also outperforms few-shot GPT-3 by a large margin on ANLI, RTE, BoolQ, AI2-ARC, OpenbookQA and StoryCloze — that is, on these datasets it beats a much larger model that was given example demonstrations in its prompt ^[1].

Ablation studies examine where the effect comes from. Three factors are identified as key to the success of instruction tuning: the number of finetuning datasets, the scale of the model, and the use of natural language instructions themselves ^[1]. Removing or varying these changes how well the model transfers to unseen tasks.

Why it matters

The result matters because it separates two things that are often bundled together: the size of a model and the way it is trained. The paper's comparison suggests that a model trained on instructions can behave more usefully on new tasks than a bigger model that was not trained this way ^[1]. It also frames zero-shot generalisation as a property that can be improved deliberately, by choosing the training data and its format, rather than as something that simply arrives with more parameters.

The framing of tasks as natural-language instructions is itself part of the contribution: it lets many different datasets be expressed in a single, shared form that a language model can read ^[1].

Versions

The paper was first submitted to arXiv on 3 September 2021 as 2109.01652 ^[1]. The current version is the fifth, revised on 8 February 2022, and the list of changes is given in Appendix F ^[1].