---
library_name: pytorch
tags:
- neuroscience
- neuronal-surrogate
- state-space-model
- time-series
---

# AxoSim model checkpoints

This repository is the model archive for [AxoSim](https://github.com/DavideWiest/axosim),
a PyTorch package for scalable neuronal-surrogate simulation. It contains every
retained AxoSim checkpoint family as of 2026-09-05. Rejected search trials,
superseded periodic snapshots, optimizer/resume state, prediction caches, and
third-party Branch-ELM weights are intentionally excluded.

## Checkpoint groups

- `released/`: the original dual-objective, structured-compact,
  regression-specialized, spike-specialized, and development-population releases.
- `ordinary/`: the four final ordinary-AxoBench publication profiles selected at
  frozen 32M/40M presentation boundaries.
- `population-finetuned/`: the same four architecture profiles after the fixed
  30K-presentation `axobench-population` specialization protocol.

`manifest.json` is authoritative: it records SHA-256, byte size, model class,
parameter count, provenance, checkpoint metadata fields, and available frozen
evaluation metrics for every file.

## Loading

Use the AxoSim source revision compatible with these checkpoints:

```bash
git clone https://github.com/DavideWiest/axosim
cd axosim
pip install -e .
```

```python
from axosim.checkpoint import load_checkpoint

model, metadata = load_checkpoint("ordinary/axosim-mamba-medium-40m.pt", map_location="cpu")
model.eval()
```

Released Mamba checkpoints require AxoSim's official fused Mamba backend. The
checkpoint embeds its architecture configuration; the JSON files in `configs/`
are included for readable reference and new training runs.

## Intended use and limits

These are research neuronal-surrogate checkpoints, not biological ground truth
or clinical models. Ordinary and population-specialized scores use different
evaluation distributions and should not be compared as if they were one test
set. Consult the AxoSim repository and its report for input/output contracts,
dataset roles, and evaluation details.

No standalone license was present in the AxoSim source repository at publication
time. The files are published for research inspection and reproducibility; ask
the AxoSim authors before redistribution or commercial use.