Marin

Marin 535B trained on 25T tokens from 152 datasets

For Marin's 535B run, the team built on 25 trillion tokens from 152 Hugging Face datasets with training-friendly licenses, sharing the pipeline work between downloading data and training.