Reproduce “Few-shot Learning with Multilingual Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal and 13 more
XGLM asks whether one multilingual model can support tasks without task-specific updates. It is trained on 500 billion tokens across 30 languages and uses task prompts, optionally with labeled examples in context, to answer without changing its parameters. On XStoryCloze, its 7.5-billion-parameter model averages 65.4% zero-shot accuracy across 11 languages, versus GPT-3 Curie 6.7B’s 54.2%; translated stories limit native-language conclusions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.