allenai/DataDecide-data-recipes

More than one training run goes into making a large language model, but developers rarely release the small models and datasets they experiment with during the development process. How do they decide what dataset to use for pretraining or which benchmarks to hill climb on? To empower open exploration of these questions, we release DataDecide—a suite of models we pretrain on 25 corpora with differing sources, deduplication, and filtering up to 100B tokens, over 14 different model sizes ranging… See the full description on the dataset page: https://huggingface.co/datasets/allenai/DataDecide-data-recipes.