Get started
Choose a dataset
Match the single-neuron, population, and intervention releases to training, selection, and evaluation questions.
AxoBench releases three datasets on Hugging Face. They share one input and target contract: native 1 ms samples, 1,278 signed event channels (639 excitatory, then 639 inhibitory), and a (time, 2) target of spike indicator and normalized somatic voltage. They differ in where the inputs come from and therefore in the question they answer.
Compare the releases
Single-neuron traces
CoreNEURON traces from five layer-5 pyramidal morphologies in train, val, and test splits. Use them for pretraining, recipe selection, and the ordinary AxoBench evaluation.
- Repository
Axym-Labs/axobench- Traces
- 125,000 × 6,000 ms
- Inputs
- A seeded event sample per trace
- Download
- 26.5 GB
Population contexts
Teacher responses to input histories recorded in 100 connected AxoSim networks, with five excitatory teachers and one inhibitory teacher. Use them to fine-tune and evaluate under recurrent input statistics.
- Repository
Axym-Labs/axobench-population- Traces
- 50,000 × 1,200 ms
- Inputs
- Recorded in connected networks
- Download
- 1.59 GB
Paired interventions
Each input sample is simulated twice: unchanged and after one of five defined edits. Use the pairs to measure response fidelity and for mechanistic analysis.
- Repository
Axym-Labs/axobench-interventions- Traces
- 12,500 pairs × 6,000 ms
- Inputs
- A baseline sample and its edit
- Download
- 5.86 GB
Bars scale each release to the largest. Simulated time is the trace count times the trace length and counts both simulations of an intervention pair; download sizes are the compressed shards of all splits.
AxoBench follows the packaging and target coordinate of NeuronIO, an earlier single-neuron corpus, which can be converted to the same shard format. These pages describe the current v1 releases.
Match the dataset to the question
Train a single-neuron model. Fit on the single-neuron train split. Select the architecture, recipe, checkpoint, and spike threshold on val, and evaluate the final choice once on test. An unreleased confirmation set is held separately from the public splits.
Compare a candidate with a baseline. Use the AxoBench evaluator on val. It scores the core metrics, calibrates the spike threshold on a separate part of the selected traces, and reports paired differences with bootstrap intervals.
Measure the response to an input change. Add the intervention pairs to the evaluation root. The evaluator then reports Response SERA, which compares the predicted change in voltage with the teacher’s change.
Prepare a model for recurrent deployment. Fine-tune and evaluate on the population contexts, splitting by context. Trace metrics on these inputs do not establish recurrent validity; connected-population measurements remain necessary.
Test transfer to an unseen morphology. No split withholds a morphology: every split contains all five. Build leave-one-morphology-out partitions from the sample IDs as described in morphologies.
Assemble an evaluation root
The evaluator reads <root>/val and, if present, every condition directory under <root>/interventions. Download only the validation split of the single-neuron release, then place the intervention release inside the same root:
python -m pip install -U huggingface_hub
hf download Axym-Labs/axobench --repo-type dataset \
--include "val/*" \
--local-dir data/axobench
hf download Axym-Labs/axobench-interventions --repo-type dataset \
--local-dir data/axobench/interventions
The first command transfers 2.65 GB of validation shards and the second 5.86 GB of intervention files. The complete single-neuron release adds 21.2 GB for train and 2.65 GB for test; the population release is 1.59 GB. Download a split with --include "train/*" or --include "test/*" when you need it.
Preserve the conventions
Single-neuron inputs use (batch, time, input_channels) and targets use (batch, time, 2). AxoSim’s population interface uses contact inputs shaped (population, time, contacts), which are a different representation from the 1,278-channel dataset inputs. Preserve channel identity, event sign, sampling cadence, morphology identity, target coordinate conversion, and causal padding rules through preprocessing. The data reference documents AxoSim’s sharded and windowed readers; the conversion reference documents the voltage and spike conversion path.
Population contexts, ordinary traces, and intervention pairs represent different input distributions. Report which release and split you use, and retain the manifest and shard hashes with derived results.
Next steps
Read the page for the selected dataset, then benchmark a model. AxoSim users can continue with training, inference-time adaptation, or evaluation.