Datasets

Select the data domain

Dataset	Use
Ordinary AxoBench	Primary single-neuron fidelity evaluation under the AxoBench metric protocol
AxoBench Population	Population-context training and evaluation
AxoBench Interventions	Matched intervention responses and mechanistic analysis

Population contexts, ordinary traces, and intervention pairs represent different distributions. Report which release and split you use, and retain the manifest and shard hashes with derived results.

Download the population and intervention releases

Install the Hugging Face CLI, then download each release separately:

python -m pip install -U huggingface_hub
hf download Axym-Labs/axobench-population \
  --repo-type dataset \
  --local-dir data/axobench-population
hf download Axym-Labs/axobench-interventions \
  --repo-type dataset \
  --local-dir data/axobench-interventions

The commands preserve each repository’s file structure under the local directory. Consult the dataset card and manifest before feeding arrays into a model; a Hugging Face checkout is not automatically a NeuronIO shard directory accepted by every trainer.

Preserve the tensor conventions

Single-neuron inputs use (batch, time, input_channels) and targets use (batch, time, 2). Population contact inputs use (population, time, contacts). Preserve channel identity, event sign, sampling cadence, morphology identity, target coordinate conversion, and causal padding rules through preprocessing.

The NeuronIO converter supports raw pickle input and deterministic shard output. The data reference documents sharded, deterministic-window, random-window, and official-style full-trace datasets; conversion reference documents the voltage and spike conversion path.

Keep splits independent

The population comparison described in the report uses 80 released training contexts, with 64 for fitting and 16 for selecting a family recipe, while the 20 released validation contexts are reserved for final evaluation. The fixed comparison budget is 30,000 presentations per model. Preserve context identity across splitting so overlapping windows from one context do not leak into final evaluation.

A new study may choose a different protocol, but it should state its context partition, recipe-selection boundary, presentation budget, and target-domain metrics. The evaluation guide describes paired ordinary-AxoBench comparisons.

Next steps

Use training to fit traces, inference-time adaptation to fit population banks, or evaluation to measure fidelity.