SparseWake

SparseWake is a synthetic benchmark for sparse temporal hydrodynamic sensing.

ICLR 2027 release

The expanded release adds controlled multi-source mixtures and common-prior nearest-source tasks, with complete core data banks, reference checkpoints, a small review supplement, and reproduction code with a frozen wake-library input.

Download release iclr2027-v1.0rc2

The version page lists the three archives, exact sizes, checksums, extraction instructions, generation scope, and component licenses. Data and result tables are CC BY 4.0; new original code is MIT; historical and third-party components retain their supplied notices. The generation code supports sensing scenes based on the cited reduced-order model, not a complete free-swimming schooling simulator or DNS reproduction.

For reproducibility, record the release identifier and final immutable repository commit. The historical root-level manifest and metadata describe the May 2026 files; the new release has its own archive checksums and package manifests.

Historical May 2026 release

The files and instructions below describe the original May 2026 release only. They are retained for compatibility; their inclusion and license statements do not override the expanded release documentation linked above.

SparseWake is an anonymous benchmark artifact for sparse temporal hydrodynamic sensing. It contains processed HDF5 datasets, metadata, summary CSV files, checksums, and lightweight Python scripts for inspecting the data and reproducing benchmark-level evaluation outputs.

Dataset Summary

SparseWake evaluates whether sparse body-fixed temporal flow measurements can recover the relative state of a neighboring wake-producing fish in a controlled single-leader setting. The benchmark data are synthetic and were generated from a DNS-parameterized fish-schooling agent-based model. The artifact starts from processed HDF5 files and does not require the upstream simulator or high-fidelity solver.

The released data support held-out-pose evaluation, temporal-history sweeps, sensor-count ablations, component-separated flow inputs, common raw-noise stress tests, sample-size convergence checks, architecture screening summaries, and paired self-signal controls.

What Is Included

• Full processed HDF5 benchmark datasets in data/processed/.
• A small sample HDF5 file in data/sample/ for inspection and smoke tests.
• Dataset metadata, schema documentation, split protocol, provenance notes, and limitations in docs/.
• Summary CSV files for reported tables and figure-level results in data/results/.
• Reproduced table CSVs in tables/.
• Lightweight loading, verification, plotting, and smoke-test training scripts in scripts/.
• Minimal reusable Python utilities in src/sparsewake/.
• Checksums in data/checksums.sha256 and a full file manifest in MANIFEST.json.
• Croissant-style metadata in croissant_metadata.json.

Main Files

Path	Purpose
data/processed/abm_sensing_dataset_v04_close_orient30_potential_160k.h5	Main external component benchmark with wake-only, potential-only, and wake-plus-potential signals.
data/processed/abm_sensing_dataset_v05_self_external_paired_160k_compressed.h5	Paired aligned-orientation self-signal control with external-only, self-only, and total-like external-plus-self arrays.
data/processed/abm_sensing_dataset_v05_self_external_orient30_160k.h5	Randomized-orientation self-signal control with external-only, self-only, and total-like arrays.
data/sample/sparsewake_sample.h5	Small sample file for quick inspection and smoke tests.
docs/data_fields.md	HDF5 fields, shapes, and label definitions.
docs/benchmark_protocol.md	Evaluation split, feature construction, standardization, and metrics.
docs/provenance.md	Data-generation provenance and non-redistributed upstream components.
dataset_card.md	Dataset-card details beyond this landing page.
evaluation_card.md	Benchmark task, metrics, ablations, and expected summary numbers.
scripts/verify_dataset.py	HDF5 structure and checksum verifier.
scripts/train_temporal_mlp.py	Small temporal-MLP smoke test and rerun utility.
scripts/reproduce_tables.py	Copies stored summary CSVs into tables/.
scripts/make_main_figures.py, scripts/make_supp_figures.py	Recreate lightweight figures from stored CSV summaries.

What Is Not Included

The upstream fish-schooling ABM source code, DNS solver files, MATLAB simulator-query code, and simulator internals are not redistributed. This release is intended to reproduce benchmark evaluations from processed HDF5 files, not to regenerate the DNS-parameterized wake library from first principles.

Intended Use

SparseWake is intended for evaluating sparse temporal physical sensing, hydrodynamic neighbor-state estimation, robustness under controlled flow-component and noise protocols, and reproducibility of benchmark-level evaluation summaries.

Out-of-Scope Use

SparseWake should not be treated as a complete biological lateral-line model, a full natural fish-schooling simulator, a pressure/shear sensing dataset, a direct CFD replacement, or a validation dataset for multi-neighbor closed-loop schooling behavior.

Data Structure

HDF5 arrays are stored in MATLAB-style or sample-first layouts depending on source export; src/sparsewake/data.py loads them sample-first. Core fields include:

• X_raw: wake-plus-potential body-frame induced-velocity features.
• X_wake_raw: wake-only component where available.
• X_potential_raw: potential-only component where available.
• X_external_raw, X_self_raw, X_total_raw: paired self-signal controls where available.
• y: labels with columns delta_x, delta_y, theta_rel, sin_phi, cos_phi, phi.
• groups: wake phase index.
• region_id: sampling region code.
• sensor_world_positions: sensor coordinates for each sample.

See docs/data_fields.md for the full schema and docs/benchmark_protocol.md for the split and feature protocol.

Quick Start

python -m venv .venv
.venv/Scripts/pip install -r requirements.txt
python scripts/verify_dataset.py --data data/sample/sparsewake_sample.h5 --checksums data/checksums.sha256
python scripts/train_temporal_mlp.py --config configs/main_v04.yaml --data data/sample/sparsewake_sample.h5 --quick
python scripts/reproduce_tables.py --results data/results --out tables

On Unix-like systems, use .venv/bin/pip and .venv/bin/python.

Reproduce Stored Tables and Diagnostic Plots

python scripts/reproduce_tables.py --results data/results --out tables
python scripts/make_main_figures.py --results data/results --out figures
python scripts/make_supp_figures.py --results data/results --out figures

These commands use stored summary CSV files and are intended for fast numerical verification and diagnostic plotting. Final manuscript figures may include manual layout adjustments and are not part of the dataset artifact.

License

The SparseWake processed datasets, metadata, result summaries, documentation, and release utilities are made available under the Creative Commons Attribution 4.0 International license (CC BY 4.0). Upstream simulator source code and solver assets are not redistributed and are not covered by this dataset release.

Citation

During anonymous review, cite this dataset as:

@dataset{sparsewake_anonymous,
  title = {SparseWake: A Temporal Hydrodynamic Sensing Benchmark},
  author = {{Anonymous Authors}},
  year = {2026}
}

Author-identifying contact information is intentionally omitted during anonymous review.