Testing, measuring and developing Welsh AI

About

An independent site observing, testing and recording the progress of AI in Welsh.

Four people, including a wheelchair user, planning together around a table

The idea in brief

Artificial intelligence models are learning to use Welsh — in text and speech — but public evidence of their ability is scattered. AIsteddfod keeps a consistent record of their progress by publishing tests, results and open research.

Why look beyond a single score?

A model may score reasonably well on a Welsh multiple-choice test yet fail to produce a valid englyn or a story that reads naturally. Form, meaning, naturalness and understanding of context need to be considered alongside scores. The Coleg Cymraeg report and our research into cynghanedd help explain those differences.

Our principles

Who is behind this?

Bangor AI’s research programme investigates strict-metre poetry and language models — from the cynghanedd rules engine to fine-tuning experiments on local clusters. The checker’s project repository is BangorAI/cynghanedd. AIsteddfod does not represent the National Eisteddfod or the Coleg Cymraeg Cenedlaethol, and has no commercial affiliation with any AI lab.

Contact

Send suggestions, corrections — especially language corrections — or ideas for new tests to aisteddfod@bangor.ai.