Marin

Marin is an open-source research platform and community for developing foundation models through transparent, reproducible experimentation. It covers the complete model-building pipeline from data curation and filtering through tokenization, pretraining, post-training, and evaluation. Experiments and decisions are documented as they occur, including unsuccessful approaches. The framework is primarily used for large language models but has also supported audio-text, DNA, and protein modeling research. Experiments are expressed as dependent steps that execute in topological order, enabling reproducible training workflows. Marin also provides model checkpoints, training recipes, scaling research, documentation, and reusable infrastructure for large-scale experiments.