Modelpublic

NathanRoll/h01-cortex-snn

sourceHugging Faceapache-2.0updated 7mo agoView on Hugging Face

1likes62downloads

Model Card

Nauro — H01 Human Cortex Connectome (Full)

The complete neuron-to-neuron connectivity matrix extracted from a nanometer-resolution reconstruction of human temporal cortex (H01 dataset, Google/Harvard/Lichtman Lab).

Built from all 166 Avro synapse shards (~32 GB raw data), filtered at ≥0.50 confidence. This is the full connectome — no spatial cropping.

Summary

Property	Value
Neurons	16,087
Excitatory	10,531 (65.4%)
Inhibitory	4,688 (29.1%)
Non-zero connections	76,903
Raw edges (pre-aggregation)	116,611
Connectivity density	0.030%
Mean in-degree	4.8
Max in-degree	70
External inputs (total)	27,022,313
Volume	Full 1 mm³
Cortical layers	L1–L6 + white matter
Build	All 166 GCS shards, min_confidence=0.50

Connectivity by cortical layer

Layer	Neurons	Exc	Inh	Internal connections	Density
Layer 1	827	85	586	55	0.008%
Layer 2	4,656	2,952	1,594	21,845	0.101%
Layer 3	2,692	1,673	965	11,018	0.152%
Layer 4	3,440	2,622	688	8,748	0.074%
Layer 5	2,313	1,665	505	6,252	0.117%
Layer 6	1,077	906	128	4,419	0.381%
White matter	648	395	111	732	0.174%

Degree distribution

Metric	In-degree	Out-degree
Mean	4.8	4.8
Std	6.3	7.0
Median	3.0	2.0
Max	70	124

Quick start

python

import json, numpy as np, torch
from safetensors.torch import load_file

# Load everything
config  = json.load(open("config.json"))
weights = load_file("connectome.safetensors")["weights"]  # (16087, 16087)
meta    = np.load("metadata.npz", allow_pickle=True)
edges   = np.load("edges.npz")["edges"]                   # (116611, 3)

print(f"{config['n_neurons']} neurons, {config['n_synapses']} connections")
print(f"Weight matrix: {weights.shape}, density: {config['density']:.4%}")

Load via HuggingFace Hub

python

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
import numpy as np

repo = "NathanRoll/h01-cortex-snn"
weights = load_file(hf_hub_download(repo, "connectome.safetensors"))["weights"]
meta = np.load(hf_hub_download(repo, "metadata.npz"), allow_pickle=True)
print(f"Loaded {weights.shape[0]} neurons")

Reconstruct from edge list

python

N = config["n_neurons"]
W = torch.zeros(N, N)
for pre, post, stype in edges:
    W[post, pre] += 1.0
# W[i, j] = number of synapses from neuron j → neuron i

Files

File	Description	Size
`connectome.safetensors`	Full 16,087×16,087 weight matrix	~1 GB
`edges.npz`	Raw edge list `[pre, post, type]`	~0.6 MB
`metadata.npz`	Positions, cell types, layers, segment IDs	~0.3 MB
`somas_filtered.csv`	Neuron table (positions, types, layers)	~1.1 MB
`config.json`	Build parameters + summary statistics	small
`layer_stats.json`	Per-layer connectivity statistics	small

Data source

The connectome data is from the H01 release by Google Research and the Lichtman Laboratory at Harvard University. The original 1.4 petabyte dataset was imaged via serial-section electron microscopy at 4 nm × 4 nm × 33 nm resolution.

Shapson-Coe, A. et al. "A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution." Science 384, eadk4858 (2024).

License

Apache 2.0. The underlying H01 data is subject to Google's release terms.