AxoBench datasets
NeuronIO compatibility
Convert NeuronIO simulations to AxoBench shards and account for the differences in metrics and channel meaning.
Shared contract
NeuronIO is an earlier single-neuron corpus of one layer-5 pyramidal cell model with 639 excitatory and 639 inhibitory synapse sites. AxoBench adopts its packaging and conventions, so a converted NeuronIO directory can be read, trained on, and scored with the same tools:
- 1,278 input channels at 1 ms, excitatory sites first with value
+1and inhibitory sites second with value-1; - targets shaped
(time, 2): a spike indicator and the somatic voltage clipped at −55 mV, shifted by +67.7 mV, and divided by 10; .npzshards withinputs,targets, andsample_ids, listed in amanifest.json.
The two corpora are not interchangeable. Their teachers, input routing, and stored arrays differ, and these differences change what the metrics measure.
Convert raw simulations
The raw NeuronIO release is distributed through Kaggle. axosim-setup --install-kaggle --download-data downloads its training and test archives to data/raw/nmda-train and data/raw/nmda-test, as described in installation. Convert a directory of raw simulation files, which follow the upstream *sim__* naming:
axobench convert-neuronio \
--input data/raw/nmda-test \
--output data/neuronio-root/val
The converter writes shards of 128 simulations named neuronio_00000.npz, with sample IDs <raw file stem>:<simulation index>. --window-size and --window-stride cut each simulation into windows with IDs ending in @<start>; with windowing, --ignore-start sets the first window start. axosim convert-neuronio performs the same conversion. The evaluator scores the shards it finds in <root>/val, so writing the converted test simulations there makes data/neuronio-root a valid evaluation root.
Differences that affect evaluation
| Property | AxoBench single-neuron traces | Converted NeuronIO |
|---|---|---|
| Teachers | Five Allen reconstructions, CoreNEURON | One layer-5 pyramidal cell model |
| Unclipped voltage | soma_mv stored with every trace |
Not stored; only the clipped target |
| Channels with a synapse | 264–370 source channels per sign, depending on morphology | Every channel corresponds to a synapse site |
| Sample-ID prefix | Morphology, specimen_<id> |
Raw file stem |
Without soma_mv, AxoBench reconstructs the teacher voltage from the clipped target. Voltage SERA, Dynamics SERA, the subthreshold RMSE, and the voltage rows of the attention panel then compare against a teacher whose action potentials are cut at −55 mV, and the action-potential windows of the subthreshold RMSE use the fixed fallback end of 4 steps. Values computed this way are not comparable with values computed against soma_mv.
The evaluator treats the text before the first colon of a sample ID as the morphology. For converted NeuronIO this is the raw file stem, so multiple_morphologies=True balances over source files, the bootstrap stratifies by source file, and the morphology-disparity row compares files rather than cells.
A model trained on NeuronIO learns a channel map in which every channel reaches a synapse. On AxoBench, a morphology-dependent subset of channels has no synapse (channel coverage). Evaluate such a model on the target corpus before drawing conclusions about transfer.
Evaluate Branch-ELM checkpoints
AxoSim’s make_official_elm_predictor loads an upstream Branch-ELM checkpoint as an AxoBench prediction function on the same target coordinate; see the prediction adapter reference. Score it with the evaluator like any other predictor.
Next steps
Use choose a dataset to decide between the corpora, or benchmark a model on the converted root.