Keynote: The Art of Language: Mastering Multilingual Challenges in LLMs
summary
Community builds an open multilingual stack: Aya dataset: 65 languages of manual instruction-style prompts, plus A collection: 500M examples of translations and templates. A volunteer network of ~3,000 people across 119 countries collects data over eight months. Models released include Aya101 (100+ languages) and Aya23 in 8B and 35B sizes, which outperform comparable multilingual baselines on several benchmarks. Evaluation suites include human-curated subsets and a Global Exams set covering ~50 languages.
AI-written summary · about
chapters (12)
- 0:00Introduction to multilingual challenges
- 2:12Multilingual representation importance
- 6:09Aya project as open science
- 12:27Aya dataset for multilingual models
- 14:57Training state-of-the-art multilingual models
- 17:49Evaluating multilingual models
- 24:54Open challenges in multilingual development
- 31:35Q&A: Unseen languages and transfer
- 33:42Q&A: Cultural biases and evaluation
- 37:17Q&A: Motivation and language nuances
- 41:40Q&A: Censorship and 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?
42mKeynote: Foundation to Build LLMs
31mKeynote Q&A – Barbara Plank
30mKeynote: LLM Research and Development at National Institute of Informatics
33mOpening Keynote: LLMs, between fantasy and reality - the poin...
15mBuilding LLMs with Cultural Context