Zyda-2 is a 5 trillion token language modeling dataset created by collecting open and high quality datasets and combining them and cross-deduplication and model-based quality filtering. Zyda-2 comprises diverse sources of web data, highly educational content, math, code, and scientific papers.
To construct Zyda-2, we took the best open-source datasets available: Zyda, FineWeb, DCLM, and Dolma. Models trained on Zyda-2 significantly outperform identical models trained on the… See the full description on the dataset page:
https://huggingface.co/datasets/jobs-git/Zyda-2.
