drooryck/multilingual-macaroni-corpus

multilingual-macaroni training corpora The 8 training corpora behind our BabyLM 2026 Multilingual-track study of code-switched pretraining curricula (English / Dutch / Chinese). All are derived from the official BabyLM 2026 multilingual corpora (babylm-{eng,nld,zho}) — no external text — and held to the same 100M byte-premium-adjusted-word budget, split across the three languages. Each corpus trains one model condition in drooryck/multilingual-macaroni-models. Subset What… See the full description on the dataset page: https://huggingface.co/datasets/drooryck/multilingual-macaroni-corpus.