AxoBench datasets

Single-neuron traces

CoreNEURON traces from five layer-5 pyramidal morphologies for training and the ordinary AxoBench evaluation.

Contents

Axym-Labs/axobench contains 125,000 traces of 6,000 ms, simulated with CoreNEURON on GPU. The teachers are five reconstructed mouse layer-5 pyramidal neurons from the Allen Cell Types Database, simulated with a Hay-compatible template and mechanisms; morphologies lists them. Every trace has its own seeded sample of excitatory and inhibitory input events, and the somatic voltage and output spikes are recorded at 1 ms.

Split Traces Per morphology NPZ shards Compressed size
train 100,000 20,000 785 21.2 GB
val 12,500 2,500 100 2.65 GB
test 12,500 2,500 100 2.65 GB

Every split contains all five morphologies. An unreleased confirmation set is held separately and is not part of the Hugging Face repository.

Shard format

Each split directory contains .npz shards of up to 128 traces, one JSON sidecar per shard, and a manifest.json that lists every shard with its trace count and sample IDs. checksums.sha256 at the repository root covers every shard and sidecar.

Array Type and shape Content
inputs int8[trace, 6000, 1278] +1 for an excitatory event on channels 0–638, -1 for an inhibitory event on channels 639–1277, otherwise 0
targets float32[trace, 6000, 2] Channel 0 is the binary spike indicator; channel 1 is the somatic voltage clipped at −55 mV, shifted by +67.7 mV, and divided by 10
soma_mv float32[trace, 6000] Unclipped somatic voltage in millivolts
sample_ids str[trace] specimen_<id>:split:<split>:<shard>:<index>:seed<seed>

The voltage target is the historical NeuronIO coordinate, (min(V, −55 mV) + 67.7 mV) / 10, so action-potential peaks are absent from targets. AxoBench scores voltage against soma_mv whenever it is present, which keeps the spike waveform in the measurement. The text before the first colon of a sample ID is the morphology identity used for balancing, stratification, and per-cell analysis.

Channel coverage

The 1,278 channels follow NeuronIO’s layout of 639 excitatory and 639 inhibitory synapse sites, and events are sampled on all of them. The five morphologies have fewer dendritic segments. Each basal or apical segment receives the source site of the same compartment type at the nearest normalized arc length, recorded as normalized_by_type_arclength in the sidecars. Source sites without a segment receive no synapse, so events on those channels do not reach the teacher.

Morphology Segments with a synapse, per sign Source channels with a synapse, per sign Source channels without a synapse, per sign
specimen_488683425 283 282 357
specimen_487664663 265 264 375
specimen_479770916 302 301 338
specimen_589427435 340 335 304
specimen_488679042 377 370 269

The values come from the input_routing record of the shard sidecars, which is identical for every trace of a morphology. A model therefore receives events on channels that carry no information for that morphology, and the informative set differs between morphologies. A morphology-conditioned model receives the identity that distinguishes the five channel maps; a model without it is trained on their mixture.

Download and load

Download one split at a time:

python -m pip install -U huggingface_hub
hf download Axym-Labs/axobench --repo-type dataset \
  --include "val/*" \
  --local-dir data/axobench

Replace val/* with train/* or test/* for the other splits. The split directories are ordinary shard directories, readable by AxoBench and by the AxoSim trainer:

from axobench.data import ShardedNeuronIODataset

val = ShardedNeuronIODataset("data/axobench/val", cache_shards=1)
inputs, targets = val.get_batch(range(4))
soma_mv = val.get_optional_array_batch("soma_mv", range(4))
print(len(val), inputs.dtype, inputs.shape, targets.shape, soma_mv.shape)
print(val.get_sample_ids([0])[0])
12500 float32 (4, 6000, 1278) (4, 6000, 2) (4, 6000)
specimen_479770916:split:val:00671:0000:seed143761753

get_batch converts inputs to float32. Reading any trace decompresses the input array of its whole shard, about 0.98 GB as int8 for 128 traces, so request indices grouped by shard and keep cache_shards small.

Use the splits

Train on train, choose the recipe and spike threshold on val, and evaluate once on test. The AxoSim training guide uses these directories directly. The AxoBench evaluator reads <root>/val; score stored predictions to evaluate the test split under the same protocol.

The AxoBench README states that the released datasets are licensed CC BY-NC 4.0: every trace is simulated on Allen Cell Types Database reconstructions, whose terms permit derived works for noncommercial purposes only.

Next steps

Check teacher statistics with the diagnostics, or add the paired interventions to the same root.

Search guides, examples, and API signatures.