Resources
A directory of corpora, test suites, models, tools and papers for everyone working on Welsh technology and language models. Send suggestions to aisteddfod@bangor.ai.
New catalogue: 104 items from Techiaith and Dewi Bryn Jones. The full list contains 47 models, 43 datasets, seven demos and seven Hugging Face collections covering voice, translation, language models and evaluation. Open the catalogue →
Text corpora
| Resource | What it is | Licence |
|---|---|---|
| CorCenCC | The National Corpus of Contemporary Welsh — over 11 million words of spoken, written and electronic Welsh | CC BY-NC-SA |
| HPLT 2.0 (cym_Latn) | Web data collected and cleaned from the Internet Archive | CC0 |
| FineWeb-2 (cym_Latn) | 96 Common Crawl snapshots in over 1,000 languages | ODC-By |
| FinePDFs (cym_Latn) | Long texts from PDFs on the web, including books and reports | ODC-By |
| Welsh Newspapers Online (National Library of Wales) | 1.1 million pages and 15 million articles from 1804–1919 | Public domain (before 1919) |
| Banc Trawsgrifiadau Bangor | Around 59,000 transcribed speech records | see the card |
| Cardiff TM en–cy | 1.45 million translation pairs | CC-BY-4.0 |
| Llyw Cymru en–cy | 980,000 bilingual pairs from the Welsh Government | OGL |
| Jupiter-N datasets (Locai) | nemotron-chat-welsh, Assembly Record, en–cy legislation, wikimedia_welsh and more | various, open |
Test suites and evaluations
| Suite | What it measures | Source |
|---|---|---|
| arc_welsh, piqa_welsh, truthfulqa_welsh | Science, common sense, truthfulness (BritEval) | UCL / BritLLM |
| llm-evals-cy | A package of ten tests: vocabulary, grammar, mutations, registers, translation… | techiaith, Bangor |
| mgsm_cy | Mathematical word problems (258 rows) | techiaith · CC-BY-SA |
| COPA-cy | Cause-and-effect reasoning (1,500 rows) | techiaith · CC-BY |
| welsh-cefr | Classifying learning materials by CEFR level (2,658 rows + two classifiers) | Cardiff |
| AIsteddfod’s cynghanedd testing protocol | Strict-metre Welsh poetry — seven syllables, valid, original and appropriate | AIsteddfod |
Models trained on Welsh data
- Jupiter-N-120B — Locai Labs/UCL’s downloadable 120B model, post-trained on Welsh. Our page →
- Caernarfon 3B — a UCL model trained entirely using computing infrastructure in Britain. Our page →
- Techiaith models — language models, speech recognition, text-to-speech and machine translation. All Techiaith and Dewi Bryn Jones models and datasets are included in our catalogue.
- BangorAI — Mistral-7B-Cymraeg, phi2-mwydryn-1 (experimental). Our page →
Language tools
- BangorAI/cynghanedd — the cynghanedd checker project repository underlying Bangor AI’s research and AIsteddfod’s cynghanedd protocol.
- Y Trawsgrifiwr (Canolfan Bedwyr) — a service converting spoken Welsh into text; Macsen — a Welsh digital assistant; Cysill/Cysgeir — language-checking tools and dictionaries; Helo Blod’s Welsh technology list.
- Techiaith speech portal — synthetic voices, text-to-speech, speech recognition and developer resources. You can also try the browser tools on our Voice.
- Bangor Pronunciation Dictionary — an open pronunciation dictionary underpinning our rules engine.
Traditional texts (public domain)
- dafyddapgwilym.net — a critical edition of 170 poems by Dafydd ap Gwilym; gutorglyn.net — Guto’r Glyn in Welsh and English; Wicidestun (cy.wikisource) — cywyddau, englynion and awdlau through the centuries.
Key papers and documents
- Jupiter-N Technical Report (Drayson, April 2026) — a method for post-training models under local control.
- Coleg Cymraeg / UWTSD report (June 2026) — AI, Welsh and Higher Education.
- Welsh Language Commissioner’s AI policy statement (July 2025) — the four regulatory principles.
- Welsh Government summary of language technology and AI 2025–26 — Teams, Copilot and grants.
- UK-LLM / BritLLM — the initiative to develop AI under British control that evaluates Welsh.
Last updated on 5 August 2026. All links were working on that date.