mysteriousauthor/spikeprophecy-steinmetz

SpikeProphecy / Steinmetz 2019 (processed)

Processed 50 ms-binned spike-count tensors derived from the publicly released Steinmetz 2019 dataset, used as the primary substrate of the SpikeProphecy benchmark (NeurIPS 2026 Datasets & Benchmarks track).

This dataset is a deterministic preprocessing of the publicly released Steinmetz 2019 recordings. The source data (raw spike times, behavioral covariates, NWB files) remain at their canonical Figshare home and are not redistributed here. We share only the binned integer-count tensors plus the session metadata that the benchmark depends on.

Files

File	Shape	Dtype	Notes
session_NNN.npy (×39)	[n_units, n_bins]	uint8	One file per session. n_units varies 228–840; n_bins is recording-length-dependent at 50 ms bin width. Counts are clipped to 255 before storage; max observed value across all sessions is 43, so no clipping occurs in practice.
metadata.json	—	—	Per-session metadata: num_units, num_bins, duration_s, split_boundaries (train_end, val_end), and per-unit brain_regions (Allen CCF acronyms). Top-level: m_max=1240, bin_width_ms=50, history_bins=10.

Provenance

Source data: Steinmetz et al. (2019), Distributed coding of choice, action and engagement across the mouse brain, Nature 576:266–273. CC-BY-4.0. https://doi.org/10.6084/m9.figshare.9598406.v2

This dataset is the output of running the SpikeProphecy preprocessing pipeline (scripts/run_ibl_cache.py-equivalent for Steinmetz NWB files; see the code repo) over the public NWB files. The pipeline is deterministic and seedless:

1. Read NWB spike times.
2. Bin at Δt = 50 ms into integer count vectors (cast to uint8).
3. Drop near-silent units below 0.1 Hz mean firing rate.
4. Emit per-session [n_units, n_bins] tensors and a 70/15/15 train/val/test temporal-split boundary in metadata.

The processing code (and a leakage-audit suite verifying the splits) is released in the SpikeProphecy benchmark code repo; see the paper for the URL.

Splits

We use a 70/15/15 train / val / test temporal split within each session: ordered first / middle / last in time. The split boundaries are encoded in metadata.json under each session's split_boundaries key as raw bin indices, not pre-sliced arrays. Random or interleaved splits would introduce trivial leakage on autoregressive forecasting because adjacent 50 ms bins are temporally correlated.

A 14-test automated leakage audit suite (released with the code) verifies five concrete leakage vectors against this split: train/test bin overlap, sliding-window boundary crossing, cross-session spillover, normalization using future statistics, and history-feature leakage.

Use

import numpy as np
import json
from pathlib import Path
from huggingface_hub import snapshot_download

local = snapshot_download(repo_id="mysteriousauthor/spikeprophecy-steinmetz",
                          repo_type="dataset")
local = Path(local)

meta = json.loads((local / "metadata.json").read_text())
print(meta["num_sessions"], "sessions, M_max =", meta["m_max"])

# Session 4 (n=703 units, used as the canonical near-median session
# in the paper's Figure 1 small multiples)
counts = np.load(local / "session_004.npy")  # [703, 60887], uint8
sb = meta["sessions"][4]["split_boundaries"]
train = counts[:, :sb["train_end"]]
val   = counts[:, sb["train_end"]:sb["val_end"]]
test  = counts[:, sb["val_end"]:]

Citation

If you use this processed dataset, please cite both the original Steinmetz paper and the SpikeProphecy benchmark:

@article{steinmetz2019distributed,
  title  = {Distributed coding of choice, action and engagement
            across the mouse brain},
  author = {Steinmetz, Nicholas A. and Zatka-Haas, Peter and
            Carandini, Matteo and Harris, Kenneth D.},
  journal = {Nature},
  volume = {576},
  pages  = {266--273},
  year   = {2019},
  doi    = {10.1038/s41586-019-1787-x}
}

@inproceedings{spikeprophecy2026,
  title  = {SpikeProphecy: A Large-Scale Benchmark for
            Autoregressive Neural Population Forecasting},
  author = {Anonymous},
  booktitle = {NeurIPS 2026 Datasets and Benchmarks Track},
  year   = {2026}
}

Croissant metadata

A Croissant 1.0 JSON-LD metadata file is included as croissant.json at the root of this dataset. It declares the file inventory, the sessions RecordSet (per-session metadata extracted from metadata.json via JSONPath), the dataset license, and Responsible-AI fields (rai:dataCollection, rai:dataPreprocessingProtocol, rai:dataAnnotationProtocol, rai:dataReleaseMaintenancePlan, rai:personalSensitiveInformation). The file validates with the mlcroissant reference implementation and supports live record iteration:

import mlcroissant as mlc
ds = mlc.Dataset(jsonld="https://huggingface.co/datasets/"
                        "mysteriousauthor/spikeprophecy-steinmetz/"
                        "resolve/main/croissant.json")
for record in ds.records("sessions"):
    print(record["sessions/index"], record["sessions/num_units"])

License

CC-BY-4.0 (matching the source dataset's license).