Keynote: The Art of Language: Mastering Multilingual Challenges in LLMs
summary
Community builds a multilingual stack: a Discord of >3,000 contributors from 119 countries collects data over eight months, producing the Aya dataset (65 manually written languages) plus an I collection of 500M examples; models include Aya101 (mt5 U base, 101 languages) and Aya23 (Cohere Command base, 23 languages, 8B/35B sizes). Claims: open models and eval suites (open‑ended, Global Exams ~50 languages) beat contemporaries on multilingual benchmarks; work includes sampling strategies, synthetic translation, and red‑teaming for safety.
AI-written summary · about
chapters (15)
- 0:00Multilingual LLMs introduction
- 1:05Multilingual representation importance
- 2:32Multilingual frameworks research
- 6:09Aya project open science
- 8:15Multilingual data collection challenges
- 12:27Aya dataset creation
- 14:57Training multilingual models
- 17:49Evaluating multilingual models
- 24:54Future directions multilingual AI
- 31:35Q&A unseen languages
- 33:42Q&A cultural biases
- 37:17Passion for multilingual apps
- 38:44Linguistic nuances biases
- 41:40Q&A model openness
- 44:05Closing remarks
related talks
47mAya: Expanding the Worlds AI serves
16mBest Practices for Open Multilingual LLM Evaluation
57mMachine Translation of Human Languages in the Age of LLMs: Is This the End of the Language Barrier?
33mOpening Keynote: LLMs, between fantasy and reality - the poin...
15mBuilding LLMs with Cultural Context
31mKeynote Q&A – Barbara Plank
42mKeynote: Foundation to Build LLMs
30mKeynote: LLM Research and Development at National Institute of Informatics