Marin
Open research platform for training large language models end to end
What it is
Marin is a Python framework and research program that spans every stage of building foundation models, from curating and filtering data and tokenizing it to pretraining, posttraining, and evaluating the resulting models. Experiments are written as dependent steps that run in order like a Makefile, and they scale from a laptop CPU to large GPU and TPU pods. ML researchers use it to reproduce and extend openly documented training runs, including the Marin 8B and 32B models and the Delphi scaling suite, and it has also been used for audio, DNA, and protein models.