README.md

---
license: other
pretty_name: AxoBench Interventions
task_categories:
  - time-series-forecasting
  - tabular-regression
tags:
  - neuroscience
  - computational-neuroscience
  - neuron-simulation
  - interventions
  - robustness
  - coreneuron
---

# AxoBench Interventions

AxoBench Interventions is the paired perturbation part of AxoBench, released
separately so intervention studies do not need to download the ordinary
single-neuron benchmark corpus. Every example contains matched baseline and
intervention inputs and teacher outputs.

## Conditions

| Condition | Paired traces | NPZ shards |
|---|---:|---:|
| `event_dropout` | 2,500 | 40 |
| `exc_dropout` | 2,500 | 40 |
| `inh_dropout` | 2,500 | 40 |
| `site_silence` | 2,500 | 40 |
| `temporal_jitter` | 2,500 | 40 |
| total | 12,500 | 200 |

Each condition directory also contains 40 JSON sidecars and one manifest.

## Shard schema

Compressed NPZ shards contain paired arrays including
`baseline_inputs`, `intervention_inputs`, `baseline_targets`, and
`intervention_targets`. The condition manifests provide sample IDs and shard
counts; `checksums.sha256` covers the complete release package.

## Loading

The Hugging Face Dataset Viewer does not natively expand these high-dimensional
NPZ tensors. Download a condition or the full repository and load shards with
NumPy:

```python
from pathlib import Path

import numpy as np

shard = next(Path("event_dropout").glob("*.npz"))
with np.load(shard, allow_pickle=False) as data:
    baseline = data["baseline_inputs"]
    intervention = data["intervention_inputs"]
```

## Related datasets

- `Axym-Labs/axobench`: ordinary isolated-neuron benchmark traces.
- `Axym-Labs/axobench-population`: connected population-context traces.

The files are provided for research use. Users are responsible for checking
the terms of the upstream simulator, morphology, and teacher-model resources
used in their application.