EM data acquisition: J. Kornfeld
Alignment: J. Kornfeld & A. Pope (V. Jain's team, Google Research)
SyConn processing: P. Schubert, A. Rother, H. Ahmad, J. Kornfeld
Cell & ultrastructure segmentation: M. Januszewski (V. Jain's team, Google Research)
EM data acquisition: J. Kornfeld
Alignment: J. Kornfeld & A. Pope (V. Jain's team, Google Research)
SyConn processing: P. Schubert, J. Kornfeld, S. Dorkenwald
Cell & ultrastructure segmentation: M. Januszewski (V. Jain's team, Google Research), S. Dorkenwald, P. Schubert
Please note that neuron reconstructions, synapse predictions, and cell-type classifications are best‑effort. We welcome feedback if you think you have discovered a systematic issue that deserves more attention.
If you use this data in your work, please cite
Rother et al., 2025
(see also songbird-connectomics.org/publications).
The full connectome is available below as a public REST API and as direct downloads of the underlying arrays, in two cell-type versions, v6 and v8 (v8 has the additional classes INT4 and OPC).
The currently available datasets, j0251 (256 × 256 × 384 µm) and j0126 (96 × 98 × 114 µm), were acquired from adult male zebra finch area X, a basal ganglia area responsible for song learning that contains striatal- and pallidal‑like cell types. Both datasets were acquired in the department of Winfried Denk at the MPI for Medical Research, Heidelberg.
The j0251 connectome (~18.36 M synapses, ~4.26 M neurons) can be queried in three ways: an interactive browser notebook, a public REST API, and direct downloads of the underlying NumPy arrays. For interactive analyses, see the API notebook; for programmatic access, the endpoints and downloads below.
The two JSON endpoints accept rich filters (cell type, probability, synapse size, post-morphology, skeleton lengths) and tolerate concurrent clients. Skeleton and mesh fetches return binary blobs. Full parameter reference is in the API notebook and the agent guide (below).
Every request names the dataset version with the full version string in the version parameter:
72_seg_20210127_agglo2_syn_20220811_celltypes_v8_20260923 (v8 cell types) or
72_seg_20210127_agglo2_syn_20220811_celltypes_20230822 (v6 cell types).
Segmentation, synapses, skeletons and meshes are the same in both; only the cell-type labels differ.
A request without a version, or with an unknown one, is answered with HTTP 400 and the list of valid versions.
Each answer carries the header X-SyConn-Version.
GET /j0251/versionscurl "https://syconn.esc.mpcdf.mpg.de/j0251/versions"
GET /j0251/celltypes?version=<version>curl "https://syconn.esc.mpcdf.mpg.de/j0251/celltypes?version=72_seg_20210127_agglo2_syn_20220811_celltypes_v8_20260923"
GET /j0251/synapses/jsonmax_results is given.
curl "https://syconn.esc.mpcdf.mpg.de/j0251/synapses/json?pre_celltype=MSN&post_celltype=GPi&return_fields=x,y,z,size,prob&max_results=100&version=72_seg_20210127_agglo2_syn_20220811_celltypes_v8_20260923"
GET /j0251/neurons/jsoncurl "https://syconn.esc.mpcdf.mpg.de/j0251/neurons/json?celltype=HVC&target_celltype=MSN&direction=outgoing&min_synapses=50&version=72_seg_20210127_agglo2_syn_20220811_celltypes_v8_20260923"
GET /j0251/skeleton?neuron_id=<id>&version=<version>curl "https://syconn.esc.mpcdf.mpg.de/j0251/skeleton?neuron_id=930416054&version=72_seg_20210127_agglo2_syn_20220811_celltypes_v8_20260923" -o skeleton.bin
GET /j0251/mesh?neuron_id=<id>&obj_type=<sv|mi|vc|sj|golgi|er>&version=<version>sv, default) or of its organelles.
Format: uint32 num_vertices, float32×3×N vertices (nm), then uint32 triangle indices
(3 per face, to the end of the data). Large cells have millions of vertices; decimate for browser rendering.
curl "https://syconn.esc.mpcdf.mpg.de/j0251/mesh?neuron_id=930416054&obj_type=mi&version=72_seg_20210127_agglo2_syn_20220811_celltypes_v8_20260923" -o mito_meshes.bin
GET /notebook/j0251/<version>/sv/<id>:0:<id>_meshcurl "https://syconn.esc.mpcdf.mpg.de/notebook/j0251/72_seg_20210127_agglo2_syn_20220811_celltypes_v8_20260923/sv/930416054:0:930416054_mesh" -o mesh.bin
Note: HVC, LMAN, and DA are axon-only in
this dataset and cannot be used as post_celltype.
Default prob_min is 0.6.
Cell-type IDs: v8 adds INT4 and OPC, so the integer cell-type IDs differ between
versions; query by name and decode returned IDs with /j0251/celltypes or the mapping files below.
Units: synapse x, y, z are dataset voxel coordinates (10 × 10 × 25 nm per voxel);
skeleton and mesh vertices are in nanometres.
The full filtered j0251 connectome as NumPy structured arrays, plus
the two label mappings. Loadable with
np.load(..., mmap_mode='r'); field schema is documented
in the agent guide.
The first four files belong to dataset version
72_seg_20210127_agglo2_syn_20220811_celltypes_20230822 (v6 cell types); the second set is the same connectome with the
v8 cell types under /connectome/data_v8/.
Synapse x, y, z are voxel coordinates (10 × 10 × 25 nm), not nanometres.
j0251_…_filtered_synapses.npy
— 613 MB, 18,361,224 synapses
j0251_…_filtered_neurons.npy
— 94 MB, 4,261,298 neurons
j0251_…_filtered_celltype_mapping.txt
— cell-type name ↔ ID
j0251_…_filtered_post_morph_mapping.txt
— post-morphology code ↔ label (0=spine_neck, 1=spine_head, 2=shaft, 3=soma)
data_v8/j0251_synapses.npy — 613 MB, 18,361,224 synapses (v8 cell-type IDs)data_v8/j0251_neurons.npy — 94 MB, 4,261,298 neurons (v8 cell types, certainty and skeleton lengths)data_v8/celltype_mapping.txt — v8 cell-type name ↔ ID (20 entries incl. INT4, OPC)data_v8/post_morph_mapping.txt, data_v8/version.txt
Cookbook-style markdown for AI coding agents (Claude and similar):
REST endpoint schemas with concrete curl examples, common filter
pitfalls, and how to load the raw .npy arrays in Python.
/connectome/connectome-guide.md