A checkpoint for exploring the batgrad training and inference workflow.

Download it with:

The checkpoint is written to outputs/checkpoints/init_baseline.pt. It contains weights for the hybrid attention/Mamba-3 model, the complete experiment configuration, training step, checkpoint format version, and creating batgrad Git commit.

Load it with batgrad.ml.checkpoint.load_checkpoint(). Mamba-3 checkpoints require Linux and CUDA. The checkpoint can be evaluated on any compatible normalized batgrad dataset. (~10M parameters, ~4GB VRAM inference)