NDT3 is a multimodal Transformer modeling neural population spiking activity from motor cortex and low-dimensional behavioral covariates. It is trained on up to 2000 hours of human and monkey neuroscience data.

We upload 45M and 350M NDT3 checkpoints for each of 200h and 2kh pretraining. These are pytorch-lightning checkpoints with model configurations wrapped into the checkpoint files. The accompanying codebase and full model card is available at: joel99/ndt3.

The configs for these models are the corresponding files on v5 experiment branch. The Weights and Biases report with pretraining curves are available here.