AxoBench datasets
Population contexts
Teacher responses to input histories recorded in connected AxoSim networks, split by network context.
Contents
Axym-Labs/axobench-population samples the input distribution that recurrent feedback produces. A thresholded AxoSim network generates branch-resolved input histories; detailed teachers of the matching cell type replay those histories to produce somatic voltage and spike labels. The labels are teacher responses to inputs recorded from the learned network, not a closed biophysical loop.
| Split | Contexts | Traces | Excitatory | Inhibitory | NPZ shards |
|---|---|---|---|---|---|
train |
80 | 40,000 | 32,000 | 8,000 | 480 |
val |
20 | 10,000 | 8,000 | 2,000 | 120 |
| total | 100 | 50,000 | 40,000 | 10,000 | 600 |
Each trace has 1,200 native 1 ms steps on the common 1,278-channel input contract. The excitatory teachers are the five morphologies of the single-neuron release; the inhibitory teacher is an Allen SST cell, allen_sst_specimen_482690728. Complete network contexts define the split, so no context contributes traces to both train and val. The release is 1.59 GB.
The release validator checked all 600 shards for schema, finite values, sample counts, role balance, manifest hashes, and disjoint context splits. This establishes package integrity; it does not show that a training recipe that improves trace metrics on these inputs also improves connected-population behavior.
Shard format
Shards are named excitatory__<morphology>__i<context>.npz or inhibitory__i<context>.npz. Each split has a manifest.json with the schema, shard counts, sample IDs, and SHA-256 hashes.
| Array | Type and shape | Content |
|---|---|---|
inputs |
int8[trace, 1200, 1278] |
Signed input events, as in the single-neuron release |
targets |
float32[trace, 1200, 2] |
Spike indicator and normalized somatic voltage |
soma_mv |
float32[trace, 1200] |
Unclipped somatic voltage in millivolts |
context_index |
int16[trace] |
Network context of the trace |
target_is_inhibitory |
bool[trace] |
Teacher role |
target_role |
str[trace] |
excitatory or inhibitory |
target_morphology |
str[trace] |
Teacher morphology identity |
window_start_ms |
int32[trace] |
Start of the 1,200 ms window |
sample_ids |
str[trace] |
<morphology>:population:<context iteration>:<target index>:<window start>:<role> |
The shards also carry target_population_index, context_iteration, context_seed, context_kind, and split. The morphology prefix of each sample ID gives six identities, so AxoBench balances and stratifies over five excitatory morphologies and the SST teacher.
Split by context
Traces from one context share its network and its activity history. Keep every partition at context granularity. The comparison in the technical report uses 64 of the 80 training contexts for fitting, the remaining 16 for selecting one learning-rate schedule per architecture family, and the 20 validation contexts only for the final evaluation, with a fixed budget of 30,000 presentations per model. This partitions the training split by context:
import numpy as np
from axobench.data import ShardedNeuronIODataset
train = ShardedNeuronIODataset("data/axobench-population/train")
contexts = train.get_optional_array_batch("context_index", range(len(train)))
rng = np.random.default_rng(0)
selection = rng.choice(np.unique(contexts), size=16, replace=False)
fit_indices = np.flatnonzero(~np.isin(contexts, selection))
select_indices = np.flatnonzero(np.isin(contexts, selection))
print(np.unique(contexts).size, fit_indices.size, select_indices.size)
80 32000 8000
A new study may choose another partition. State the context partition, recipe-selection boundary, presentation budget, and target-domain metrics with the result.
Download and evaluate
hf download Axym-Labs/axobench-population --repo-type dataset \
--local-dir data/axobench-population
The validation split has the directory layout that the AxoBench evaluator reads, so evaluate_model(predict, dataset_root="data/axobench-population", multiple_morphologies=True) scores it under the same metric protocol. The default 500-step exclusion leaves 700 scored steps per trace; state it, or set EvaluationConfig.ignore_start, when you report results. In single-morphology mode the evaluator selects the alphabetically first identity, which here is the inhibitory SST teacher; pass morphology_id or multiple_morphologies=True deliberately. The release contains no intervention pairs, so Response SERA is unavailable.
Trace fidelity is not recurrent validity
In the technical report, training on all 50,000 traces improved held-out Voltage SERA, Dynamics SERA, and spike F1, yet the connected model missed the morphology-conditioned firing-rate criterion on all three unseen connectomes. Use population-context metrics to select a model for this input distribution, and connected-population measurements to decide whether it behaves correctly in recurrent simulation.
Next steps
Fine-tune a trained neuron model on these traces as part of inference-time adaptation, then simulate it in a connected population.