Dataset Catalog¶

The datasets of the Seahorse benchmark are hosted on the seahorse-stpp Hugging Face organization and load by a short name. They share one trait — events observed in space and time — but the meaning of "space" stretches from a city block to the human cortex. The catalog is ordered along that arc.

Every dataset loads the same way, by the short name shown on each card. The repositories are pinned to their release revision and every split is verified against the hashes on the dataset card:

from seahorse.data import load_dataset, load_released_dataset

splits = load_dataset("covid")                       # {"train": [...], "val": [...], "test": [...]}
splits = load_released_dataset("austin311")          # same, with hash verification and provenance
native = load_released_dataset("austin311", config="native")

…or from the command line with --dataset covid. The full seahorse-stpp/<id> reference still works everywhere a short name does. Cards marked paper are the ten real-data corpora and the synthetic controls of the Seahorse benchmark; legacy marks the three Neural-STPP benchmark splits kept for comparison with earlier work; not benchmarked marks released datasets outside the paper. Three repositories are script-only: their source licences do not permit hosting the data, so they ship a rebuild script and the expected hashes instead.

Moving through the city¶

space · docks and pickup points

NYC bike share · 2019paperscript-only citibike_nyc Bike-share ride starts across New York City in 2019; rebuilt locally from the operator's trip files. spacedock coordinatestimeride start Uber pickups · NYCpaper uber_nyc Ride-hailing pickup requests across New York City (2014). spacepickup locationtimerequest time NYC bike share · legacy splitlegacyscript-only citibike_legacy The Neural-STPP bike-share benchmark (times in days), reconstructed by a rebuild script. spacedock coordinatestimeride start US traffic accidentsnot benchmarked us_accidents Road traffic accident reports across the United States. spaceaccident locationtimereport time

Incidents & city services¶

space · where something was reported

Chicago crimepaper chicago_crime Reported crime incidents across the city of Chicago. spaceblock locationtimeincident time LA crimepaper la_crime Reported crime incidents across Los Angeles. spacereport locationtimeincident time NYC motor-vehicle collisionspaper nyc_collisions Police-reported motor-vehicle collisions across New York City. spacecrash locationtimecrash time Seattle fire 911paperscript-only seattle_fire911 Fire-department 911 dispatches in Seattle; rebuilt locally from the city's open-data export. spacedispatch addresstimedispatch time Austin 311paper austin311 Non-emergency city service requests in Austin, Texas. spacerequest locationtimerequest time NYC 311paper nyc311 Non-emergency service requests across New York City. spacerequest locationtimerequest time

Earth & environment¶

space · latitude and longitude on the globe

ComCat · Californiapaper comcat_california USGS ANSS ComCat earthquake catalog for California, native time stamps. spaceepicentertimeorigin time GeoNet · New Zealandpaper geonet_nz GeoNet earthquake catalog for New Zealand. spaceepicentertimeorigin time Earthquakes · legacy splitlegacy earthquakes The Neural-STPP earthquake benchmark split (USGS); some test windows overlap training windows in time. spaceepicentertimeorigin time US wildfiresnot benchmarked us_wildfires Wildfire ignition records across the United States. spaceignition locationtimediscovery date

Population health¶

space · case geography

COVID-19 · legacy splitlegacy covid Reported COVID-19 cases over space and time (The New York Times); the Neural-STPP split, whose validation and test windows lie inside the training period. spacecase locationtimereport date

Beyond the map¶

space · not a map at all

BOLD5000non-2Dnot benchmarked bold5000 Eventized neural responses derived from BOLD5000 fMRI recordings — included to show STPP space need not be geographic. spacethe braintimescan time

Supported — but not in the benchmark

BOLD5000, US wildfires and US traffic accidents are released and load like every other dataset. The paper's comparison covers the ten real-data paper datasets above, the synthetic controls, and the three legacy Neural-STPP splits, under one evaluation protocol.

Synthetic controls¶

space · whatever you configure it to be

Seahorse synthetic controlspapersynthetic synthetic Three generators with analytic ground truth — homogeneous_poisson, moving_seasonal_poisson, gamma_renewal — selected with config=. spaceunit squaretimeknown process

When you need known ground truth — to isolate one factor that real data confounds — Seahorse also uses synthetic sequences from HawkesNest.

HawkesNest · entanglement suitesynthetic github.com/YahyaAalaila/HawkesNest Spatio-temporal sequences generated with tunable entanglement between the space and time dimensions, under known ground truth — so a model's behaviour can be probed against a factor that real-world data cannot cleanly separate. spaceconfigurable domaintimeconfigurable process

Want to add your own? See Add Your Dataset for the preparation checklist and Conversion Standard for the JSONL format.