rigidhat's picture
Upload README.md with huggingface_hub
b00b800 verified
|
Raw
History Blame Contribute Delete
2.6 kB
metadata
license: mit
language:
  - en
tags:
  - construction-safety
  - osha
  - regulatory-compliance
  - low-resource
  - niche-domain
size_categories:
  - 10K<n<100K

Construction Code-Citation Corpus v1

Open dataset of construction-site incident narratives paired with OIICS hazard codes (event, source, nature, body) and OSHA 29 CFR 1926 citation candidates. Built for the Adaption Labs AutoScientist Challenge ("All Other Domains" category).

Sources

  • OSHA Severe Injury Reports (DOL, public domain): 2015-01 → 2025-08, 103,750 records. Each row has Final Narrative (incident text) plus OIICS classification codes for the event, source, nature, and body part.
  • OSHA 29 CFR 1926 (eCFR snapshot 2025-09-16, public domain): 304 sections with explicit citations and full regulatory text.

Schema

Each row in the canonical jsonl:

{
  "id": "<SIR ID>",
  "input": "<Final Narrative>",
  "hazards": [{
    "code_event":  {"id": "<OIICS event>",  "title": "..."},
    "code_source": {"id": "<OIICS source>", "title": "..."},
    "code_nature": {"id": "<OIICS nature>", "title": "..."},
    "code_body":   {"id": "<OIICS body>",   "title": "..."},
    "severity": "low|moderate|high"
  }],
  "citations": [],
  "naics": "<6-digit NAICS>",
  "naics_subsector": "<4-digit>",
  "event_date": "YYYY-MM-DD",
  "inspection_nr": "<int or null>",
  "source": "sir",
  "split": "train|dev|test"
}

The citations field is empty in v1 (SIR does not carry OSHA standard citations directly). Citation supervision comes from a separate join on inspection_nr to the DOL OSHA enforcement violations corpus, planned for v2.

Splits

Stratified by NAICS subsector (first 4 digits), 70/15/15.

  • train: 72,467
  • dev: 15,410
  • test: 15,873 (SHA-256 c9490ed3..., hash-pinned, never re-shuffled)

Known biases

  • SIR over-represents severe injuries (hospitalization, amputation, loss of eye) — the corpus is by definition skewed toward high-severity events.
  • Source-code distribution has a heavy long tail: 1,478 unique codes, top-75 cover only 57% of records. Models will need either a code-collapse strategy or hierarchical (division-level) prediction.

License

MIT (this dataset). OSHA SIR is public domain. OSHA 29 CFR 1926 text is public domain.

Citation

If you use this dataset, please cite:

@misc{construction-code-corpus-2026,
  title  = {Construction Code-Citation Corpus v1},
  author = {Oversite Innovations},
  year   = {2026},
  note   = {Built for the Adaption Labs AutoScientist Challenge}
}