MSHA enforcement citations

Predict whether a US mine safety inspection will cite a given 30 CFR section - one binary task instance per section, over 26 years of federal enforcement records.

About this release

MSHA publishes its violations, inspections and mine records; anyone can download them. The contribution here is not access. It is the frozen panel of 497,897 inspections, the declaration that the prediction instant is the moment the visit begins, the prior-history features computed only from that mine's strictly-earlier inspections, the time-ordered partitions, and a leakage gate that held 45 per-section tasks to one contract.

The task is binary and per-section: will this inspection cite 30 CFR 56.12004? The label is the set of sections the inspector actually wrote, taken from the violation rows, so this is an enforcement outcome - what the inspector found - and not a hazard measurement. Each of the 45 shipped sections is a complete task instance over the identical pool: train 2001-2020, eval 2021-2023, holdout 2024-2025.

The numbers are published with their floor. On the reference section sec_56_12004 the baseline reaches 0.8036 eval AUC, identical from the training path and an independent validation re-run. The persistence reference - the same model using only whether that section was cited at the mine's previous inspection - reaches 0.5831, so the full feature set clears it by 0.2205. Section prevalence ranges around 9-11% on the reference task; score AUC, not accuracy.

Splits

Qualification and baseline

PASSED - 45/45 per-section configs, 32 checks each, 0 failed, on the full artifact on the worker Gate 1.3.0.

One baseline through the runner's training and validation contract. No agent or harness comparison was run and none is implied.

Use this version

from datasets import load_dataset
ds = load_dataset('earino/msha-enforcement-citations', 'sec_47_41_A', revision='v2026.10')
# 45 configs, one per target: substitute any other name from README.md

The version links and checksums identify this artifact. Exclude the labelled holdout from the agent's workspace when running an evaluation.

Files and checksums

Artifact eff3c196ff7bda848991a4bb8bbf8c4e061d6f956217cfcfa54935ad8c96bda3

Limitations

• A derived research dataset from MSHA public-domain data, labelled as derived - not an official MSHA or Department of Labor product.
• The label is an enforcement outcome: what an inspector cited, not a measured hazard. Mines that are inspected more often generate more citations.
• Only sections cited in at least 1% of pooled inspections ship as configs, so this is not a census of 30 CFR.
• A mine is the same entity across splits by design - prior history is the point of the task. The temporal split is what prevents leakage, and the holdout is labelled and downloadable: keep its labels outside an evaluated agent's workspace.

Source and terms

MSHA Open Government Data Sets: Violations, Inspections and Mines (Mine Safety and Health Administration, US Department of Labor)

Mine Safety and Health Administration (MSHA), US Department of Labor, MSHA Violations Data Set (and the companion Inspections and Mines data sets), https://arlweb.msha.gov/OpenGovernmentData/DataSets/, accessed 2026-10-12.

Source: LicenseRef-Public-Domain. Compilation: CC0-1.0. Code and documentation: MIT.

Our contribution: the row selection, the derived multi-label target (the set of 30 CFR sections cited at one inspection), the declared prediction instant, the temporal partitions, the feature construction and the packaging. The dedication covers our contribution only. It does not cover or relicense the source rows: those are published by MSHA without terms and are redistributed here as US federal government works.

Suggested citation

Mine Safety and Health Administration, US Department of Labor. MSHA Violations Data Set. US federal government work, public domain by statute (17 USC 105). Accessed 2026-10-12. Derived dataset: MSHA enforcement citations, inspection-level multi-label release, version to be set at packaging.

Versions

• v2026.10