earino/msha-enforcement-citations

MSHA Enforcement Citations (v2026.10)

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

Will a US mine safety inspection cite a given 30 CFR section? One binary task instance per section, 45 of them, over the same 497,897 pooled inspections and the same time-ordered splits. Every instance carries 25 columns: an event id, the mine and inspection attributes recorded before or at the visit, five prior-behaviour features from that mine's strictly-earlier inspections, a per-section persistence feature, and the label.

The source is the Mine Safety and Health Administration's published enforcement record: violations, inspections and mines, 2000 onward, covering coal and metal/non-metal operations. The task asks what an inspector was about to find, using only what was knowable when the visit began.

The task

For config sec_56_12004, the target is "this inspection will cite 30 CFR 56.12004". The label is 1 when the inspection's violation rows include that exact PART_SECTION value, 0 otherwise. The prediction instant is the inspection's INSPECTION_BEGIN_DT; nothing recorded after the visit begins is a feature.

Rows and splits are identical in every config, by design — the configs are alternative binary targets over one panel, not different populations:

• train 2001–2020 — 415,088 inspections
• eval 2021–2023 — 49,233 inspections
• holdout 2024–2025 — 33,576 inspections

Rows: 497,897 inspections. Total across the three splits; every config carries the same rows. Split boundaries are calendar dates, so the eval and holdout windows are later in time than training; no inspection appears in more than one split.

Each config also ships meta.json (feature list, label definition, provenance, split windows) and quality.json (per-split missingness, cardinality and prevalence). The extract root carries labels.csv: one row per pooled inspection, one column per shipped section, so the joint multi-label problem can be reconstructed from the same bytes.

Choosing a config

Each config is a complete, independently scoreable binary task with its own train/eval/holdout files. sec_56_12004 is the calibrated reference point: the shipped baseline scores 0.8036 eval AUC on it. Prevalence varies by section and by window (0.0886 / 0.1063 / 0.0996 on the reference config), so compare a model across configs by AUC, not accuracy.

Baseline

One baseline was measured on sec_56_12004 through the runner's own training and validation contract, on the exact bytes in this release (fetched from staging, not rebuilt):

• eval AUC 0.8036, identical from the training path and an independent validation re-run; train AUC 0.8036; CONTRACT OK
• persistence reference — the same model using only prev_target_cited — 0.5831, so the 24-feature model clears the section's own persistence by 0.2205

Scope of these numbers: one baseline through the runner's contract. No agent or harness comparison was run and none is implied.

Loading

The repository is public, so loading needs no credential:

from datasets import load_dataset

ds = load_dataset("earino/msha-enforcement-citations", "sec_56_12004", revision="v2026.10")
# every config is loaded the same way; the config must be named

The GitHub counterpart carries the identical files and digests: https://github.com/earino/msha-enforcement-citations, with MANIFEST.json, SHA256SUMS and get_dataset.py.

Provenance and licence

Source data: MSHA, US Department of Labor — public domain by statute (17 USC 105); the publisher states no terms. 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.

Our contribution — row selection, the derived section targets, the declared prediction instant, the temporal partitions, the features and this packaging — is dedicated to the public domain under CC0-1.0. Code is MIT. The dedication covers our contribution only and does not relicense the source rows.

Limitations

• The label is the set of sections the inspector cited, which is an enforcement outcome, not an underlying hazard measurement. Mines with more inspections generate more citations.
• Prior-behaviour features are computed from the mine's earlier inspections only, but a mine is the same entity across splits by design — the task is about a mine's future, so entity overlap is the point, and the temporal split is what prevents leakage.
• Sections cited in fewer than 1% of pooled inspections are not shipped as configs, so this is not a census of the 30 CFR part.
• No agent or harness comparison was run on this artifact and none is implied.

Dataset Factory · v2026.10

Dataset website · GitHub release · Hugging Face · Catalogue

Artifact: eff3c196ff7bda848991a4bb8bbf8c4e061d6f956217cfcfa54935ad8c96bda3