AxoBench datasets
Morphologies
The five excitatory teacher morphologies, the inhibitory population teacher, and leave-one-morphology-out partitions.
Excitatory teachers
The single-neuron and intervention releases use five dendrite-only reconstructions of mouse layer-5 spiny neurons from the Allen Cell Types Database. The population release uses the same five as its excitatory teachers. specimen_488683425 is the reconstruction bundled with AxoBench as the template; the other four were ranked by a compatibility score against it, where lower is closer.
| Morphology | Cre line | Role | Branches | Total length | Radial extent | Apical fraction | Score |
|---|---|---|---|---|---|---|---|
specimen_488683425 |
Rbp4-Cre_KL100 | reference_control |
99 | 1.000 | 1.000 | 0.000 | 0.000 |
specimen_487664663 |
Rbp4-Cre_KL100 | closest_heldout_control |
98 | 0.930 | 1.022 | +0.053 | 0.203 |
specimen_479770916 |
Rorb-IRES2-Cre | intermediate_line_shift |
110 | 1.116 | 0.930 | −0.029 | 0.247 |
specimen_589427435 |
Sim1-Cre_KJ18 | line_diversity_shift |
120 | 1.187 | 0.975 | +0.054 | 0.284 |
specimen_488679042 |
Rorb-IRES2-Cre | stronger_sane_length_shift |
132 | 1.325 | 0.986 | +0.019 | 0.405 |
Total length and radial extent are ratios to the reference, and the apical fraction is a difference from it. The score adds the absolute logarithms of the total-length, routed-length, and radial-extent ratios, the absolute apical-fraction difference, and the branch-count difference of the routed representation divided by its 45 branches. The values come from the dataset manifest and axobench.benchmark.v1_morphologies(); axobench benchmark-v1 writes them as routing QC rows.
The morphologies also differ in how many of the 639 source channels per sign reach a synapse; see channel coverage.
Inhibitory teacher
The population release adds one inhibitory teacher, the Allen SST cell allen_sst_specimen_482690728. It supplies all 10,000 inhibitory population traces and appears in no other release.
Morphology identity in the data
The text before the first colon of a sample ID identifies the morphology, and shard file names begin with it. AxoBench uses this prefix in three places:
evaluate_model(..., morphology_id="specimen_488679042")evaluates one morphology. Withoutmorphology_id, single-morphology mode uses the alphabetically first identity,specimen_479770916.multiple_morphologies=Trueinterleaves the morphologies, so the default 120 traces contain 24 of each and the calibration block contains all five.- Paired bootstrap intervals resample traces within each morphology, and the subthreshold RMSE fixes its action-potential window per morphology.
AxoSim models that condition on morphology need the same identity as an index into their declared morphology vocabulary. The AxoSim prediction adapter maps sample-ID prefixes to that index and rejects identities outside the vocabulary, so a held-out morphology must be declared before the model can be evaluated on it.
Hold out a morphology
No released split withholds a morphology: train, val, and test each contain all five. To test transfer, build a leave-one-morphology-out partition. The technical report uses five such comparisons: each source model is trained on the other four morphologies, and target adaptation and evaluation use separate traces of the omitted one.
Every shard contains a single morphology, so a source directory can link the shards of four morphologies together with the original manifest. AxoBench and the AxoSim trainer read only the manifest entries whose shards are present:
held_out=specimen_488679042
mkdir -p data/lomo-$held_out/train
ln -s "$PWD/data/axobench/train/manifest.json" data/lomo-$held_out/train/
for shard in data/axobench/train/specimen_*.npz; do
case "$shard" in
*"$held_out"_*) ;;
*) ln -s "$PWD/$shard" data/lomo-$held_out/train/ ;;
esac
done
The source directory then holds 80,000 traces. Adapt to the held-out morphology on its training traces, select on its validation traces, and report the final result on its test traces, for example with evaluate_model(predict, dataset_root="data/axobench", morphology_id=held_out) during selection. Selecting the adaptation recipe on the evaluation traces would make the comparison optimistic.
Compare morphologies
The attention row morphology_error_disparity reports the worst morphology’s voltage RMSE relative to the median morphology whenever an evaluation contains at least two. For a fuller comparison, run the evaluator once per morphology_id with the same sample limit, or derive morphology_contrast rows from per-morphology diagnostic rows as described in diagnostics.
Next steps
Use a held-out partition with inference-time adaptation, or benchmark each morphology separately.