Nethobench/nethobench-widefield-v1

Nethobench Widefield Calcium Forecasting Dataset

Summary

This release packages a benchmark-ready widefield calcium imaging dataset for neural time-series forecasting and Nethobench-style evaluation.

The release contains two complementary data representations:

1. A prepared benchmark tensor: data/data100_ba16.npy, already organized into fixed-length subsequences for model training and evaluation.
2. An unprepared source Parquet table: data/data-clean-all.parquet, containing the unstandardized, unprepared continuous sequences before conversion into the benchmark tensor.

The primary benchmark signal is stored as a single NumPy array with Parquet manifests for Hugging Face dataset viewer preview and Croissant tooling.

Files

• data/data100_ba16.npy: prepared float32 benchmark tensor, shape [sequence, subsequence, region, time].
• data/data100_ba16_metadata.json: original sidecar metadata from preprocessing, also mirrored under metadata/source_metadata_raw.json.
• data/data-clean-all.parquet: unstandardized, unprepared source data table. This file contains the continuous wide-format neural traces before organization into fixed-length subsequences. It is useful for users who want to inspect or rebuild the benchmark tensor from the less processed sequence-level data.
• metadata/: corrected release metadata, protocol JSON, and preprocessing notes.
• manifests/: Parquet and CSV tables describing subsequences, short windows, and valid long-rollout starts.
• scripts/: load_dataset.py, make_examples.py, validate_release.py.

Source Parquet format

The file:

data/data-clean-all.parquet

contains the unstandardized, unprepared sequence-level data used upstream of the benchmark tensor construction.

It is a wide-format table with sequence/time indexing columns and neural region columns. Conceptually, rows correspond to timepoints within original recording sequences.

Expected indexing columns:

sequenceId
itemPosition

The remaining columns correspond to neural activity traces for brain regions. Unlike data100_ba16.npy, this file is not already arranged as [sequence, subsequence, region, time], and it is not the direct model input format used in the benchmark protocol.

Use this file if you want to:

• inspect the less processed continuous sequences;
• rebuild the prepared tensor with a different subsequence length;
• choose different brain regions or sequence subsets;
• audit the preprocessing pipeline.

For benchmark training/evaluation, use:

data/data100_ba16.npy

and the associated manifest/protocol files.

Tensor format

The prepared benchmark tensor is stored in:

data/data100_ba16.npy

with shape:

array.shape = [287, 211, 16, 180]
axes        = [sequence, subsequence, region, time]

The same shape is duplicated in metadata/release_metadata.json under array_shape.

Brain regions

The release uses 16 mesoscale regions. Region names:

• primary-motor-area-layer-1-a
• primary-motor-area-layer-1-b
• primary-somatosensory-area-barrel-field-layer-1-a
• primary-somatosensory-area-barrel-field-layer-1-b
• primary-somatosensory-area-lower-limb-layer-1-a
• primary-somatosensory-area-lower-limb-layer-1-b
• primary-somatosensory-area-upper-limb-layer-1-a
• primary-somatosensory-area-upper-limb-layer-1-b
• primary-visual-area-layer-1-a
• primary-visual-area-layer-1-b
• retrosplenial-area-dorsal-part-layer-1-a
• retrosplenial-area-dorsal-part-layer-1-b
• retrosplenial-area-lateral-agranular-part-layer-1-a
• retrosplenial-area-lateral-agranular-part-layer-1-b
• secondary-motor-area-layer-1-a
• secondary-motor-area-layer-1-b

Short-window training protocol

For each stored subsequence chunk, whose local shape is [region, time]:

x = array[seq, sub, :, :90].T      # shape [90, 16]
y = array[seq, sub, :, 90:180].T   # shape [90, 16]

Model convention: input [time, region], i.e. time-major after the transpose.

Long-rollout evaluation protocol

The model is trained on 90 → 90 short windows. For long benchmarks, it is rolled out 720 steps into the future using recursive multi-step prediction in downstream code. Ground truth for scoring uses 810 consecutive timesteps: 90 context timesteps plus 720 future target timesteps.

Because each prepared subsequence stores only 180 timesteps, the 720-step future must be stitched from the tail of the starting subsequence and subsequent subsequences within the same sequence, in chronological order.

No-padding long examples require at least four full subsequences after the initial partial tail. Equivalently, valid starting subsequence indices run from:

0 through n_subsequences_per_sequence - 5

For this release, with 211 subsequences per sequence, this means:

0 through 206

Example usage:

from pathlib import Path

from scripts.load_dataset import load_array, get_short_xy, get_long_rollout_xy

root = Path(".")  # if your current working directory is this release folder
arr = load_array(root)

x_short, y_short = get_short_xy(arr, 0, 0)
x_long, y_long, gt_long = get_long_rollout_xy(arr, 0, 0)

print(x_short.shape, y_short.shape)
print(x_long.shape, y_long.shape, gt_long.shape)

Expected shapes:

x_short: [90, 16]
y_short: [90, 16]

x_long:  [90, 16]
y_long:  [720, 16]
gt_long: [810, 16]

Alternatively, run modules from the scripts/ folder:

python scripts/load_dataset.py --root .

Splits

The tensor is released unsplit at the file level.

For convenience, manifests include a deterministic sequence-level split using:

seed = 101
train_ratio = 0.8

This split is implemented to match split_by_sequences in preprocess_helpers.py, using numpy.random.seed and permutation.

Always report the exact split used in publications.

Normalization

Values are released as stored.

The training code in this repository normalizes after the train/validation split using training input timesteps only. This avoids:

• leakage across train/validation splits;
• leakage from target/future timesteps into normalization statistics.

The source Parquet file data/data-clean-all.parquet is provided as an unstandardized, unprepared sequence-level table. The prepared tensor data/data100_ba16.npy is the benchmark-ready format used by the supplied loaders and manifests.

Intended use

This dataset is intended for:

• benchmarking neural time-series forecasting models;
• evaluating synthetic or predicted neural dynamics with Nethobench-style scorers;
• studying short-window and long-rollout behavior of forecasting models;
• reproducing the benchmark protocol described in the accompanying Nethobench paper.

Limitations

• Mouse widefield calcium imaging; mesoscale summaries, not spikes.
• Calcium signals are indirect proxies for neural activity and are temporally slower than electrophysiological recordings.
• Limited number of recording sequences compared to large-scale generic ML datasets.
• Long-rollout construction assumes subsequences are chronological and contiguous within each sequence.
• The prepared tensor reflects one specific region selection, subsequence length, and preprocessing path.
• The source Parquet file is provided for transparency and reproducibility, but the benchmark protocol is defined on the prepared tensor.
• Not for clinical or diagnostic use.

Citation

For anonymous review, citation details are withheld. The final citation will be added after acceptance.