> Instruction tuning enables pretrained language models to perform new tasks from inference-time natural language descriptions. These approaches rely on vast amounts of human supervision in the form of crowdsourced datasets or user interactions. In this work, we introduce Unnatural Instructions: a large dataset of creative and diverse instructions, collected with virtually no human labor. We collect 64,000 examples by prompting a language model with three seed examples of instructions and eliciting a fourth. This set is then expanded by prompting the model to rephrase each instruction, creating a total of approximately 240,000 examples of instructions, inputs, and outputs. Experiments show that despite containing a fair amount of noise, training on Unnatural Instructions rivals the effectiveness of training on open-source manually-curated datasets, surpassing the performance of models such as T0++ and Tk-Instruct across various benchmarks. These results demonstrate the potential of model-generated data as a cost-effective alternative to crowdsourcing for dataset expansion and diversification.

Found out about it from https://dblalock.substack.com/i/93113359/unnatural-instructi..., which gives the following surprising implication:

> Importantly, their results also seem to get better when the models generating the example inputs and outputs get more sophisticated. This suggests a virtuous cycle of smarter language models yielding better training data yielding smarter language models.

> Between this and the previous paper, it feels like language models creating their own data is about to be a mainstream thing. Since high-quality data being expensive has been so fundamental to ML and statistics for so long, I’m not even sure what all the implications of this will be…