Skip to content

Dataset Overview

Seahorse reads event sequence data from JSONL split files — one JSON object per line, one line per sequence. The same format works for single-model runs, benchmark campaigns, and Hugging Face sources.

What a Dataset Needs

A complete dataset for a Seahorse run consists of three split files:

File Required for Notes
train.jsonl fit, tune, bench Training sequences
val.jsonl fit, tune, bench Validation sequences — drives early stopping
test.jsonl evaluation Held-out test sequences

Each file is one JSON object per line. Each object is one event sequence.

Minimal Record

{"times": [0.1, 0.4, 1.2], "locations": [[0.2, 0.4], [0.3, 0.8], [0.7, 0.1]]}

times[i] and locations[i] describe the same event. locations must be a list of 2D coordinate pairs.

Data Sources

Source How to use
Released seahorse-stpp datasets load_dataset("covid") or --dataset covid (short name, pinned revision)
Any Hugging Face repo Pass --dataset owner/repo[/subdir] — Seahorse downloads and caches
Local JSONL files Pass --train, --val, --test flags or a directory with --dataset
Local split collection Point --splits_dir at a root with one subdirectory per dataset

See Data Format for the full contract and command support matrix.

Ready-to-use Datasets

The datasets of the Seahorse benchmark are released in the seahorse-stpp Hugging Face organization — urban service and incident records, bike-share and ride-hailing demand, earthquake catalogs, public health, synthetic controls, and a few supported datasets outside the paper — all in the same JSONL split format. Load any of them by its short name; every repository is pinned to its release revision and each split is verified against the hashes on the dataset card:

from seahorse.data import load_dataset, load_released_dataset

splits = load_dataset("covid")                   # downloads + caches
splits = load_released_dataset("austin311")      # same, with hash verification

Two datasets (citibike_nyc, seattle_fire911) cannot be hosted under their source licences; their repositories ship a rebuild script and the expected hashes instead.

Browse the full collection, with load snippets and each dataset's space/time axes, in the Dataset Catalog.

Need controlled ground truth? Seahorse's synthetic benchmark sequences are generated with HawkesNest — its entanglement suite produces spatio-temporal data with tunable space–time coupling for stress-testing models. See Synthetic controls.

Next Steps

Was this page helpful? Glad it helped — a ⭐ star on GitHub helps other researchers find Seahorse, and here is how to cite it. Sorry that missed — tell us what's missing.
Suggest edits Raise issue