ConnectomeBench2

ConnectomeBench2 is a unified benchmark for automated proofreading of connectomic neural-segmentation data. 716,485 samples across 4 species (mouse, fly, human, zebrafish) and 5 sample types (real merge edits, real split edits, synthetic adjacent / junction / synapse controls), with the associated mesh geometry and electron-microscopy (EM) renderings.

Downstream trainers should treat this dataset as the single source of truth for sample identity, labels, train/validation/test split, and which task(s) a row is valid for.

v2 (June 2026) — major schema refresh. The prior v1-may06 release (401,170 rows) is preserved as a git tag; load it via load_dataset("jeffbbrown2/ConnectomeBench2", revision="v1-may06"). Breaking changes vs v1 are summarized in "Changelog" at the bottom.

Context: Connectomic Proofreading

Connectomics scans and automatically segments neurons to create large-scale brain maps at cellular resolution. Two types of segmentation errors can occur in this process, which need to be corrected (= proofreading):

• False Splits — corrected via merge corrections
• False Merges — corrected via split corrections

Merge corrections (of false splits) are applied to multiple segments that need to be correctly merged together. Split corrections (of false merges) are applied to single segments that need to be correctly split apart.

For this reason, this dataset contains renderings of both single-segment (pre-split or post-merge) and dual-segment (post-split or pre-merge) mesh geometry. Each row carries one geometry render whose dual-vs-single semantics is determined by the row's sample_type. EM data is provided in dual format only — segmentation on imaging level is contiguous, so the single-version can be derived from the union of the dual.

Renderings (geometry and EM imaging data)

(top: synapse merge-pair — both masks populated; bottom: junction control — single-mask only, mask B / seg B empty)

Geometry files (the geometry column) are compressed .npz payloads that decode to (3, 7, 224, 224) float16 arrays — three 2D views (front, side, top) × seven channels:

ch	content
0	silhouette
1	depth
2	normal_x
3	normal_y
4	normal_z
5	mask A
6	mask B (empty in single-segment renders)

The geometry column is dual-segment when sample_type ∈ {merge_edit, adjacent_control, synapse_control} and single-segment when sample_type ∈ {split_edit, junction_control}. The two flavors differ not only in mask channels but also subtly in all other channels, due to slight differences in mesh geometry from merging/splitting. See the per-row metadata.is_merge field (or equivalently, sample_type) to disambiguate.

Free split-mask labels. For split_edit rows, three additional PNG columns — split_mask_front, split_mask_side, split_mask_top — provide per-view, pixel-level ground truth for the post-split (dual) state, view-aligned with the single-segment geometry render. Split-mask-generation tasks consume these directly without extra labeling. Coverage: 127,445 / 145,338 split_edits (87.7%); the rest had no after-state available.

EM imaging files (em_xy_before / em_xz_before / em_yz_before / em_best_before columns) are PNG-encoded 3-channel slices of the before-edit state:

ch	content
0	raw EM intensity
1	segment A mask
2	segment B mask

Four imaging views per sample: three cardinal slices (xy, xz, yz) + a best slice at an oblique angle that maximizes the visible area of both segments (sum of their logs). The _before suffix is to be explicit that these reflect the pre-edit segmentation; no _after EM is rendered.

For single-segment tasks, segment A and B should be merged (and B zeroed). The best view may leak some dual-label information (it takes both labels into account); we advise against testing single-segment tasks on em_best_before.

Loading

from datasets import load_dataset

ds = load_dataset("jeffbbrown2/ConnectomeBench2", split="train")
sample = ds[0]
# sample["em_xy_before"] is a PIL Image (HF auto-decodes)
# sample["geometry"] is bytes — decode with:
import io, numpy as np
geom = np.load(io.BytesIO(sample["geometry"]))["arr_0"]   # shape (3, 7, 224, 224) float16

Or with raw pyarrow:

import pyarrow.parquet as pq
import numpy as np, io
df = pq.read_table("train/train-00000.parquet").to_pandas()
geom = np.load(io.BytesIO(df.iloc[0]["geometry"]))["arr_0"]

The metadata/{train,val,test}.parquet sidecars contain identifier/label/modality columns only (no image bytes) — useful for fast filtering or inspection.

Columns

Identifiers

• combined_sample_hash — primary key (md5 hex 32-char of f"{source_archive}|{source_archive_sample_hash}"); guaranteed unique across the dataset.
• source_archive_sample_hash — legacy 12-char hex hash from upstream; kept for traceability, not unique alone.
• source_archive — name of the originating render archive (e.g. unified_mouse, unified_controls_fly). 8 distinct values (unified_{sp} for ops/adj, unified_controls_{sp} for junction+synapse, × 4 species).

Sample identity

• sample_type: str — single source of truth for what kind of sample this row is. Five values:
  • merge_edit — positive merge-correction edit
  • split_edit — positive split-correction edit
  • adjacent_control — synthetic negative for merge-correction (segments adjacent to genuine correction)
  • junction_control — putative junction in proofread neuron (negative merge-error-id sample)
  • synapse_control — synapse pair across neurons (negative merge-correction)
• same_neuron: bool — derived from sample_type:
  • True for merge_edit, junction_control
  • False for split_edit, adjacent_control, synapse_control
• species: str — fly / mouse / human / zebrafish.

Image content

• geometry — bytes; compressed npz (key "arr_0") decoding to (3, 7, 224, 224) float16. Always present in v2. Dual-segment when sample_type ∈ {merge_edit, adj_ctrl, syn_ctrl}, single-segment when sample_type ∈ {split_edit, junction_ctrl}.
• em_xy_before / em_xz_before / em_yz_before / em_best_before — PIL Images (3-channel PNG, (224, 224, 3) uint8). Always present in v2.
• split_mask_front / split_mask_side / split_mask_top — PIL Images, per-view after-state split GT for split_edit rows. Null for other sample types. ~87.7% coverage on split_edits.
• has_em: bool — true if any em_*_before column is non-null. True for every row in v2.
• has_after_mask: bool — true iff the three split_mask_* columns are populated. Only ever true for sample_type == split_edit.
• present_slots: list[str] — modality tags actually present (e.g. ["em_best_before", "em_xy_before", "em_xz_before", "em_yz_before", "geometry"] or with + "split_mask_*" for split_edits with after-mask).

Task routing & labels

• task_routing: list[str] — which downstream task(s) this row can serve as training data for. Computed from sample_type + has_after_mask:
  • false_split_correction — merge-correction task. Fires for sample_type ∈ {merge_edit, synapse_control, adjacent_control}.
  • false_merge_identification — merge-error binary classification. Fires for sample_type ∈ {split_edit, junction_control}.
  • split_mask_generation — pixel-level split prediction. Fires for sample_type == split_edit AND has_after_mask.
• false_split_correction_label: int | null — 1 for merge_edit, 0 for synapse_control / adjacent_control, null for split_edit / junction_control. Trainers filter by task_routing (or check for non-null label).
• false_merge_identification_label: int | null — 1 for split_edit, 0 for junction_control, null for the other three. Same filtering rule.

Usage note. Downstream training scripts must load the appropriate geometry flavor per task:

• Merge Correction of false splits should use dual-segment geometry (rows where sample_type ∈ {merge_edit, adj_ctrl, syn_ctrl})
• Split Correction of false merges should use single-segment geometry (rows where sample_type ∈ {split_edit, junction_ctrl})
  • Furthermore, fuse A/B channels of EM images and discard em_best_before (it sees both labels at oblique angle and can leak ground truth)

Otherwise, ground-truth task or label information may leak to the model and bias performance.

Train/val/test split

• split: str — train / validation / test. Target ratios 75/12.5/12.5; observed ~74.2/11.6/14.2 (slight hash-based per-cube assignment noise at the scale of one volume). Assigned by spatial location of the proofreading sample — edit_point_nm for operations and adjacent controls, interface_point_nm for junction controls, synapse_ctr_pt_nm for synapse controls — bucketed into 80µm cubes. Cube extent is the canonical segmentation-volume bbox per species (queried from CloudVolume), not the min/max of observed bank points.

Other

• metadata: str — JSON-stringified original metadata struct. Parse with json.loads. Useful keys: operation_id, source_operation_id, strategy, image_types, interface_point_nm, edit_point_nm, before_root_ids, after_root_ids, is_merge, species, …

Counts

• 716,485 rows total · ~74/12/14 train (531,734) / validation (82,822) / test (101,929)
• All rows have geometry + 4 EM views; 127,445 split_edits additionally have 3 split-mask views
• ~5.1M model-level samples counting per-modality views (3 geom views + 4 EM views) × 716,485 + 3 split-mask views × 127,445
• 903 parquet shards (~250 MB each) — 669 train / 105 val / 129 test

Layout

README.md
shards.csv                    metadata across shards (path, sha256, n_samples, size)
train/train-*.parquet         WebDataset-style parquet shards with image bytes
val/val-*.parquet
test/test-*.parquet
metadata/                     sidecar parquets with identifiers + labels (no bytes)
  train.parquet
  val.parquet
  test.parquet
demo.parquet                  stratified mini-shard (one-line preview)
figures/
  channel_decomposition.png

Sources & License

Derived from the following upstream connectomic proofreading datasets:

• MICrONS (mouse cortex)
• FlyWire (Drosophila brain)
• H01 (human cortex)
• Zebrafish larval connectome

License = other; users must comply with upstream licenses (which may differ across species/sources). Final outbound license will be set after upstream license review.

Citation

If you use ConnectomeBench2, please cite:

Brown, J., Farkas, T., Razgar, G., Boyden, E. S.
ConnectomeBench2: A unified benchmark for automated connectomic proofreading.
(2026, in submission). Brown J. and Farkas T. contributed equally as first authors.

Please also cite the upstream connectome sources used by this dataset:

• MICrONS (mouse cortex): https://www.microns-explorer.org/cortical-mm3
• FlyWire (Drosophila): https://flywire.ai/
• H01 (human cortex): https://h01-release.storage.googleapis.com/landing.html
• Zebrafish (fish1): https://fish1-release.storage.googleapis.com/index.html

Changelog

v2 (June 2026)

• Row count: 401,170 → 716,485
• EM coverage: now 100% across all sample types (was partial in v1)
• geometry_single column removed: only geometry exists now; its dual-vs-single semantics derive from sample_type / metadata.is_merge
• EM column renamed: em_{xy,xz,yz,best} → em_{xy,xz,yz,best}_before (no _after EM exists; suffix makes the before-edit state explicit)
• Split-mask columns added: split_mask_{front,side,top} PNG per-view labels for split_edit rows
• Flags reworked: has_single_mask, has_dual_mask → has_after_mask. Use metadata.is_merge (or sample_type) to distinguish dual-vs-single geometry render.
• Label semantics nullable: false_split_correction_label / false_merge_identification_label are now non-null only for relevant sample types (was always populated in v1, derived from same_neuron)
• Split assignment: 80/10/10 → 75/12.5/12.5; 50µm → 80µm cubes; bbox now from CloudVolume (canonical) instead of bank min/max; per-sample-type coord choice (edit_point_nm for ops/adj, interface_point_nm for junction_ctrl, synapse_ctr_pt_nm for synapse_ctrl)
• Source archives: 10 → 8 (unified_{sp} and unified_controls_{sp} × 4 species)

To load the old version: load_dataset("jeffbbrown2/ConnectomeBench2", revision="v1-may06").