The training data for drooryck/babylm-macaroni, the BabyLM 2026 Multilingual-track submission. All corpora are created from the official BabyLM 2026 multilingual corpora (babylm-{eng,nld,zho}) and embedded with code-switching using an instruction-tuned LLM.

	The four corpora

Folder Description

shuffled_cs/ Full code-switched corpus, all documents globally shuffled.

shuffled_nocs/ Matched unilingual twin (same documents… See the full description on the dataset page: https://huggingface.co/datasets/drooryck/babylm-macaroni-corpus.