---
library_name: pytorch
datasets:
- marplan6/batgrad
tags:
- battery
- timeseries
- foundation-models
- template
license: mit
---

# batgrad model

A checkpoint for exploring the
[`batgrad`](https://github.com/marplan/batgrad) training and inference workflow.

Download it with:

```sh
uv run scripts/hf_assets.py download --ckpt init_baseline
```

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)