---
license: other
pretty_name: AxoBench
task_categories:
  - time-series-forecasting
  - tabular-regression
tags:
  - neuroscience
  - computational-neuroscience
  - neuron-simulation
  - single-cell
  - benchmark
  - dendrites
  - coreneuron
---

# AxoBench

AxoBench is a large-scale CoreNEURON corpus for training and evaluating
data-driven models of layer-5 pyramidal neurons.
It contains 125,000 six-second traces across multiple reconstructed
morphologies, with event inputs and paired somatic-voltage/spike targets.

Two companion datasets cover distinct evaluation settings:

- [`Axym-Labs/axobench-population`](https://huggingface.co/datasets/Axym-Labs/axobench-population)
  contains population-compatible shards and connectivity context.
- [`Axym-Labs/axobench-interventions`](https://huggingface.co/datasets/Axym-Labs/axobench-interventions)
  contains paired causal interventions.

## Layout

- `train/`: 785 compressed NPZ shards, 100,000 traces.
- `val/`: 100 compressed NPZ shards, 12,500 traces.
- `test/`: 100 compressed NPZ shards, 12,500 traces.

Each compressed NPZ shard contains arrays including `inputs`, `targets`, and
`sample_ids`; the adjacent JSON file records shard-level provenance and
metadata.

## Compression

The payload is stored as shard-wise deflated NPZ files, matching the practical
NeuronIO-style packaging while preserving partial downloads and resumable
uploads. Do not wrap this folder in one monolithic zip or tar archive for the
primary Hugging Face distribution: the shards are already compressed and an
outer archive would make subset access worse.

See `checksums.sha256` for file integrity checks.

## Metadata

- `dataset_package_manifest.json`: split counts, sizes, and packaging details.
- `checksums.sha256`: SHA-256 checksums for all trace and shard-metadata files.

## Access State

The public `test` split supports reproducible comparison.
An unreleased confirmation set is held separately and is not part of this Hub
repository.