earino/austin-911-response

Austin 911 Response Time (2026.09)

Status: public. Released after explicit operator authorisation; publication changed no data file, tag or checksum.

Predict whether an Austin 911 call for service will have a late first-unit arrival: at least 1200 seconds between the call being answered and the first unit arriving. The label and every feature are knowable at prediction time; the label source column is never shipped.

• Artifact version: e4598317e406984fa590aacc5e7aff675867578c51ae1a84ebf6279bc42c3328
• Version tag: v2026.09 (use this revision for reproducibility)
• GitHub counterpart: https://github.com/earino/austin-911-response
• Rows: 1,049,636 across three disjoint temporal splits

Loading

Version-pinned. revision is the tag, not a branch, so the bytes cannot change under you.

from datasets import load_dataset

ds = load_dataset("earino/austin-911-response", revision="v2026.09")
print(ds)
# DatasetDict({
#     train: Dataset({features: [...], num_rows: 572180}),
#     eval: Dataset({features: [...], num_rows: 285665}),
#     holdout: Dataset({features: [...], num_rows: 191791})
# })

# the label is `late`, a binary target
print(ds["train"].features["late"])

The repository is public, so loading needs no credential and no token has to be passed:

ds = load_dataset("earino/austin-911-response", revision="v2026.09")

The files are also plain CSVs, so they can be used without datasets:

hf download earino/austin-911-response --repo-type dataset --revision v2026.09 \
    --include "data/*.csv" --local-dir ./austin

Splits

Split by response_datetime, disjoint and ordered, so no split leaks into another:

split	rows	positives	positive rate	window
train	572,180	231,025	0.4038	before 2025-01-01
eval	285,665	109,860	0.3846	2025
holdout	191,791	72,450	0.3778	2026 onward

The labels

late is the binary target: 1 when the first unit arrived 1200 or more seconds after the call was answered.

The threshold was frozen from the training window by rule, not by hand - the largest tested value whose training-window positive rate stays at or above 0.40, counting only aggregate queries:

• 1200s gives a training rate of 0.403763
• 0.390050 falls below the target: 0.390050 falls below the 0.40 target

Positive rates differ by split (0.4038 / 0.3846 / 0.3778); that reflects when the data was collected, not a construction fault.

The response-time clock (measured, not assumed)

response_time is documented as measuring from when the call was answered, and no column records that instant, so the label cannot be recomputed from the shipped fields. Measured over 2000 rows:

• definition: (first_unit_arrived_datetime - response_datetime) - response_time
• median -57.0s, p05 -206.0s, p95 0.0s, range -468.0s to 6041.0s
• 1890 negative / 55 exact zero / 55 positive rows

Reproduce it with python3 code/measure_clock_offset.py 2000 --json.

Provenance

• Source: Austin Open Data - APD 911 Calls for Service 2023-2026 (dataset e687-fx2y), https://datahub.austintexas.gov/d/e687-fx2y
• Licence of the source data: Public Domain, per the publisher's catalogue metadata.
• Extraction: 69 HTTP requests (15 aggregate, 54 pages), measured 353.7s on 2 vCPU / 2 GB worker.
• Rebuild: MANIFEST.json pins the build timestamp, so a rebuild is byte-identical.

Licence and attribution

Source data (City of Austin)	Public Domain, preserved unchanged - designation preserved, attribution and citation given below
Our rights in this derived compilation	CC0-1.0 (LICENSE-CC0-1.0.txt)
Code and documentation in the GitHub counterpart	MIT (LICENSE-MIT.txt)

Scope, stated plainly: the CC0 dedication covers our contribution - the row selection, the derived binary label, the temporal partition and the packaging. It does not cover or relicense the City of Austin's data, which keeps its own Public Domain designation. This dataset is derived from the City of Austin's open data and is labelled as derived; it is not an official City of Austin product.

Attribution. City of Austin, APD 911 Calls for Service 2023-2026, Austin Open Data, https://datahub.austintexas.gov/d/e687-fx2y, accessed 2026-09-18

Suggested citation.

City of Austin. APD 911 Calls for Service 2023-2026. Austin Open Data. Public Domain. Accessed 2026-09-18. Derived dataset: Austin 911 response time prediction dataset, version 2026.09, artifact e4598317e406984fa590aacc5e7aff675867578c51ae1a84ebf6279bc42c3328.

Future datasets in this series choose terms compatible with each source; CC0 here is not a standing permission to override another source's terms.

Qualification and baseline

• Gate: PASSED - 33 checks, 0 failed, on the full 1,049,636-row artifact, on the worker (gate 1.3.0), run over the full artifact on a worker.
• One baseline through the runner's existing training and validation contract: eval AUC not recorded.
• Scope: One baseline through the runner's training and validation contract. No agent or harness comparison was run and none is implied.

Every number above is measured, with the command that produced it recorded in measurements.json; figures that are estimates are labelled as such.

Dataset Factory · v2026.09

Dataset website · GitHub release · Hugging Face · Catalogue

Artifact: e4598317e406984fa590aacc5e7aff675867578c51ae1a84ebf6279bc42c3328