haidang2405/aqshift-us-leakage-safe-air-quality
AQShift-US AQShift-US is a leakage-resistant benchmark built from real U.S. EPA Air Quality System measurements. It supports next-local-day forecasting and interval calibration under forward-time and unseen-site shift for daily maximum 8-hour ozone and observed 24-hour PM2.5. The frozen source snapshot covers 2010–2025 and was published by EPA AirData on 2026-06-25. It contains 7,544,381 mature benchmark examples derived from 6,234,916 canonical ozone monitor-days and 1,447,795… See the full description on the dataset page: https://huggingface.co/datasets/haidang2405/aqshift-us-leakage-safe-air-quality.
AQShift-US
AQShift-US is a leakage-resistant benchmark built from real U.S. EPA Air Quality System measurements. It supports next-local-day forecasting and interval calibration under forward-time and unseen-site shift for daily maximum 8-hour ozone and observed 24-hour PM2.5.
The frozen source snapshot covers 2010–2025 and was published by EPA AirData on 2026-06-25. It contains 7,544,381 mature benchmark examples derived from 6,234,916 canonical ozone monitor-days and 1,447,795 canonical PM2.5 monitor-days. Calendar-day lags, reason-coded target validity, deterministic split roles, source-member hashes, and revision warnings are explicit.
This is a retrospective measurement benchmark. It is not a real-time air quality service, health-risk model, regulatory determination, or EPA product.
Load a table
Each logical table is an independent Hugging Face subset because the tables have different schemas. The default subset is benchmark_examples.
from datasets import load_dataset
examples = load_dataset(
"haidang2405/aqshift-us-leakage-safe-air-quality",
"benchmark_examples",
split="data",
streaming=True,
)
first = next(iter(examples))Use canonical_daily, monitors, site_roles, or sites as the second argument to load another table. Column semantics are documented in `release/DATA_DICTIONARY.md`.
Integrity and reproducibility
The directory release/ is a byte-identical copy of the audited public release. Download the repository and run:
python release/verify_release.py --bundle releaseExpected release SHA-256: dee1eccb05f313245f19af19a924a8332f034a61458b73070358b44fb19712d4. The verifier checks 14 manifest-bound files and 360,142,290 bytes.
The full scientific data card, target construction, leakage boundary, split design, known limitations, baseline metrics, source-object manifest, and changelog are preserved inside release/.
Source, rights, and attribution
Source measurements are U.S. EPA Air Quality System (AQS) ambient monitoring data obtained from official AirData pre-generated files. EPA states that these monitoring data are public domain. Authored documentation, split definitions, manifests, and derived compilation are offered under CC BY 4.0; original code is Apache-2.0. See `release/LICENSE.md` for the exact license map and attribution.
AQShift-US is independent and is not endorsed, approved, or certified by EPA. Source measurements are not real-time and may be revised.
