BEPAD — Benchmark Earthquake Phase Association Dataset

Thousands of phase-association problems built from real earthquakes and real picks of UTDQuake. Each window holds several overlapping earthquakes on one seismic network, with real, generated and noise picks, and the answer (earthquake) for every pick.

One window (Texas network): (a) earthquakes and stations, (b) every pick at its distance from its earthquake, (c) picks by type. The associator's job is to find which picks belong together.

v0.1 is built from UTDQuake 1.0 (Hugging Face commit 8524c6f0): every seismic network, named by its UTDQuake code (tx, CM, R08 ...). A station is NET.STA (FDSN network and station code).

Quick start

import bepad                                   # pip install bepad
test = bepad.load("test", networks=["tx"], max_windows=20, labels=False)   # 20 Texas test windows
for w in test:
    w.picks, w.stations                        # the inputs of one association problem
report = bepad.evaluate(predictions, test)     # score your associator

Without the package:

import pandas as pd
picks = pd.read_parquet("hf://datasets/ecastillot/BEPAD/v0.1/test/picks/network=tx.parquet")

Content

Folder	What
v0.1/<split>/windows	one row per window: network, scenario, counts, difficulty, seed
v0.1/<split>/events	the earthquakes of each window (origin, location, magnitude, scorable)
v0.1/stations.parquet	station coordinates; station = NET.STA, network = UTDQuake network
v0.1/<split>/window_stations	the stations available in each window (window_id, station); coordinates in stations.parquet
v0.1/<split>/picks	one row per pick: station, phase, time; labels event_idx (−1 noise), pick_type (real / generated / noise), real, prob (optional confidence for training, from detectability)
v0.1/metadata/	window statistics as CSV
v0.1/label_model.json	predicted arrivals, timing uncertainty and prob model
pool/	the quality-controlled earthquakes and picks (for generating new training windows)
associators/<name>/	contributed associators (code + parameters), runnable with bepad run
leaderboard/	submitted results (scores + predictions) and the tables in leaderboard/README.md

Per network and scenario: 1,000 train, 200 val and 100 test windows; 23 scenarios (1–20 earthquakes per window × 5 noise levels, plus earthquakes outside the network). Every earthquake in a window has ≥ 4 real P picks on ≥ 3 stations inside it.

Split	Windows	Networks
train	2,068,000	90
val	412,800	90
test	208,900	91

Networks with fewer than 200 earthquakes are test only (ESPOO, R06).

BEPAD-10

The main ranking uses the BEPAD-10 test: the test windows of 10 networks chosen for different scales, geometries and depths, from a 4 km fracking array to national networks: R05 ToC2ME, R07 LASSO, R01 COSEISMIQ, nc N. California, tx Texas, JMA Japan, CM Colombia, R09 N. Chile (IPOC), ak Alaska, NOA Greece. Every leaderboard entry must run them:

bepad run associators/my-method --split test --networks core --out results/my-method
bepad submit results/my-method --code associators/my-method --dry-run    # your row, nothing is sent

Leaderboard

See leaderboard/README.md: the BEPAD-10 table first (mean over its networks), then all networks, then one table per network. To add a method: bepad new, bepad run, bepad submit — the last command opens a pull request on this repository that adds only your two folders.

Benchmark rules

1. Report on the test split. 2. Use only picks (without label columns) and stations as input.
2. Treat each window separately. 4. Tune only on train / val; velocity models must be published or fitted on train / val.
3. Run at least the BEPAD-10 networks. 6. Submit your code with your results (bepad submit --code): the test labels are public, so entries are re-run before they are merged.

License and citation

Data follow the terms of the original providers as distributed by UTDQuake; please cite UTDQuake and the network operators. Code (the bepad package): GPL-3.0.