GYAFC

GYAFC, short for Grammarly Yahoo Answers Formality Corpus, is a benchmark dataset used in natural language processing for formality style transfer. It provides a parallel corpus of informal and formal sentence variants to support training and evaluating models that convert text from informal to formal register and vice versa.

The dataset is built from Yahoo Answers posts and curated rewrites produced by Grammarly editors to reflect formal language. It is organized into two domains drawn from Yahoo Answers categories: Entertainment & Music and Family & Relationships. Each entry consists of an informal sentence paired with a corresponding formal rewrite, enabling supervised learning for style transfer.

GYAFC has been widely adopted as a standard resource for researching formality transfer because it offers real-world, domain-specific informal language paired with professionally edited formal equivalents. Researchers use it to develop and assess neural and non-neural style-transfer methods, including sequence-to-sequence and transformer-based approaches, and to compare performance using automatic metrics alongside human judgments of formality.

The dataset is publicly available for academic use and is accompanied by splits for training, validation, and testing, as well as documentation on its creation and quality control procedures. While it provides a valuable benchmark, users should consider domain biases and the subjective nature of formality when interpreting results. GYAFC is often discussed in conjunction with the Yahoo Answers Formality Corpus (YAFC), on which it builds.