wtf is gpt-3?

This repository is the companion to OpenAI's research paper introducing GPT-3, an AI language model trained on 175 billion parameters, at the time ten times larger than any comparable model. The paper, published in 2020, demonstrated that a sufficiently large language model can learn to perform new tasks from just a handful of written examples, without needing to be retrained on labeled data for each task specifically. This ability is called "few-shot learning." The repository itself does not contain GPT-3's code or weights. Instead, it hosts materials that accompany the paper: sample text outputs generated by the model, synthetic datasets used for certain word-scramble and arithmetic tests, statistics about the training data, and a model card (a summary document describing what the model does and its limitations). You would visit this repository if you are reading the GPT-3 research paper and want access to the datasets or sample outputs it references, or if you need to cite the paper in your own work. It is primarily useful to researchers and students studying the history of large language models. There is nothing to install or run here, it is a research artifact archive, not a deployable tool.