Self-Instruct: language models taught to follow instructions using synthetic self-generated data

Yizhong Wang and colleagues introduced Self-Instruct, a method for training language models to follow instructions using synthetic data the models generate themselves, closing much of the gap with InstructGPT-001 on evaluated tasks while using far less human annotation.

Background

Instruction-tuned language models, ones fine-tuned to respond to requests rather than just predict the next word, had shown a striking ability to handle tasks they had never seen during training. The catch was data. Building that kind of model meant collecting large amounts of text written by humans who had carefully composed instructions and correct responses. That data was scarce, and what existed tended to be narrow.

The problem was not just quantity. Human-written instruction datasets were also limited in diversity and creativity, which meant models trained on them inherited those limits. Getting better coverage meant paying for more annotation, and there was no obvious ceiling on how much you might need.

What happened

Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi and Hannaneh Hajishirzi introduced Self-Instruct in a paper submitted to arXiv on 20 December 2022. The core idea was to let the language model generate its own training material.

The pipeline works in a loop. The model produces candidate instructions, then generates corresponding inputs and outputs for each one. A filtering step removes examples that are invalid or too similar to ones already collected. What remains goes back in as fine-tuning data for the same model. The process requires only a small fixed set of human-written seed examples and is otherwise automatic.

Applied to the standard GPT-3 without any other modifications, Self-Instruct produced a 33% absolute improvement over the base model on Super-NaturalInstructions, a large benchmark of instruction-following tasks. That put it on par with InstructGPT-001, which had been trained using private user data and human annotations that Self-Instruct did not require.

The team also put together a set of expert-written instructions covering novel tasks and ran human evaluations against it. GPT-3 tuned with Self-Instruct outperformed models tuned on existing public instruction datasets by a large margin, finishing only 5% behind InstructGPT-001 in absolute terms.

The code and the synthetic dataset were released publicly, giving other researchers a foundation to build instruction-tuning work without having to start from scratch on data collection.

Why it mattered

Before Self-Instruct, aligning a language model to follow instructions required large quantities of human-written data, which was expensive and limited in variety. The method showed that a model could bootstrap its own training signal, closing a substantial portion of the gap with InstructGPT-001 on the evaluated human-preference task while requiring only a small fixed set of human-written seed examples and otherwise running automatically. The authors released their synthetic dataset to support further work on instruction tuning.

Sources

• Self-Instruct: Aligning Language Models with Self-Generated Instructions. arxiv.org. Primary source
• Self-Instruct: Aligning Language Models with Self-Generated Instructions. arxiv.org. Primary source

Cite this page

AI Achievements. (2022). Self-Instruct: language models taught to follow instructions using synthetic self-generated data. Retrieved 2026-08-29, from https://achievements.ai/milestone/self-instruct-language-models-taught-to-follow-instructions

@misc{achievements_self_instruct_language_models_taught_to_follow_instructions,
  title  = {Self-Instruct: language models taught to follow instructions using synthetic self-generated data},
  author = {{AI Achievements}},
  year   = {2022},
  url    = {https://achievements.ai/milestone/self-instruct-language-models-taught-to-follow-instructions}
}