ComplianceKB / README.md
eugeneasun's picture
Publish versioned ComplianceKB dataset and v5 research card
51cbe93 verified
|
Raw History Blame Contribute Delete
6.63 kB
metadata
pretty_name: ComplianceKB - Source-linked OSHA regulatory requirements
language:
  - en
license: cc-by-4.0
size_categories:
  - n<1K
task_categories:
  - text-retrieval
tags:
  - osha
  - compliance
  - regulatory-knowledge
  - occupational-safety
  - human-ai-verification
configs:
  - config_name: benchmark_v1_2_0
    default: true
    data_files:
      - split: train
        path: requirements_v1_2_0.jsonl
  - config_name: extended_v1_5_0
    data_files:
      - split: train
        path: requirements_v1_5_0.jsonl

ComplianceKB

Source-linked regulatory knowledge for human-supervised compliance AI. Research by Eugene Sun and Dupeng Liu, conducted during a TrusOne internship.

GitHub project · Current v5 A0 poster · v5 methods, outputs and reproduction · Original manuscript

Two separate snapshots

Configuration Records Meaning
benchmark_v1_2_0 (default) 108 Frozen corpus used in the original study and current v5 development evaluation
extended_v1_5_0 133 Later coverage extension, not evaluated by the reported 96 v5 calls

These snapshots overlap; do not add their row counts and describe 241 distinct requirements. They are version alternatives, not independent train/test partitions. train is the Hub file split label only. No independent holdout is supplied.

Contents and sources

Selected requirements from four U.S. federal OSHA frameworks: Process Safety Management (29 CFR1910.119), Hazard Communication (1910.1200), injury/illness recordkeeping (Part 1904), and hazardous-energy control/LOTO (1910.147). Coverage is selective, not exhaustive OSHA coverage.

Each record links a requirement restatement to modality, risk, descriptive applicability, a workflow trigger, evidence category, verification method and official source excerpt/clause/URL/version/hash. Trigger and evidence annotations make proposed checks inspectable and reusable; they are not proof of actual workplace performance.

The viewer JSONL flattens applicability/provenance and normalizes list fields. Curator/review-person identifiers are omitted from the viewer. Original, already public release JSON files are retained byte-identically as original_v1.2.0.json and original_v1.5.0.json. Field names retain source semantics; list normalization does not add legal interpretations. Inspect those original snapshots for complete provenance and optional fields.

Curation and verification

Official-source extraction/AI-assisted drafting was followed by author source review and correction, with frozen releases. The 108-record experimental corpus received author review with AI assistance. Fifty retained IDs from the original 89-record pool received blinded second-person labels. Unadjudicated nominal Cohen kappa: risk .234, evidence .642, trigger .443. Agreement measures consistency, not expert legal truth; it does not certify the extended corpus.

Current presentation: v5

Local framework routing and query-term expansion -> BM25 up to 20 candidate records -> excerpts or excerpts plus RKM -> one call to one selected provider. No preliminary LLM call or dual-model voting. The 24 scenarios are AI-assisted, author-confirmed, repeatedly inspected development cases. Both representations share candidate IDs. Medium effort,6,000-token output cap. Exact recorded IDs: gpt-5.6-terra / claude-sonnet-5.

Model Input Req F1 Trigger F1 Evidence F1 Decision
GPT-5.6-terra BM25 excerpts 0.894 0.669 0.559 0.958
GPT-5.6-terra BM25 excerpts + RKM 0.912 0.986 0.940 0.958
Claude Sonnet 5 BM25 excerpts 0.876 0.436 0.633 0.958
Claude Sonnet 5 BM25 excerpts + RKM 0.859 0.888 0.890 1.000

96 saved answers (24 x 2 inputs x 2 models) are available on GitHub. Trigger/evidence scores are model-generated labels. Citations are linked from selected IDs. RKM improves trigger/evidence recovery in both models, while requirement gains vary. GPT decision accuracy ties; Claude improves by 1/24. Operational reference labels share the curated taxonomy, and added RKM fields add information/tokens. Scores do not establish independent legal correctness or all-metric superiority.

Use

from datasets import load_dataset
records = load_dataset("eugeneasun/ComplianceKB", "benchmark_v1_2_0", split="train")
extended = load_dataset("eugeneasun/ComplianceKB", "extended_v1_5_0", split="train")
print(records[0]["source_clause"], records[0]["workflow_trigger"])

No custom dataset loading script or model API is required. See GitHub for code and offline reproduction. Intended uses include regulatory retrieval research, source-linked audit-check proposals, annotation studies and human-AI review. This is a dataset repository, not trained model weights or a deployed compliance product.

Limitations

  • Stored excerpts and source versions are dated snapshots, not a real-time legal service. Hashes detect stored-text drift, not legal validity or temporal reasoning.
  • Applicability is descriptive, not an executable/legal-complete predicate. Scope, exceptions, technical classifications and workplace facts require qualified verification.
  • Evidence categories describe what to inspect; they do not establish legal documentation duties or authenticate workplace records.
  • No independent legal-expert test set, calibrated confidence, measured human-time savings or production safety certification is claimed.
  • Synthetic development scenarios are not evaluated real facilities. Underlying pretrained models may know public OSHA text.
  • Do not use this resource alone to authorize hazardous work or make final compliance judgments.

License and citation

Curated data/annotations: CC BY 4.0, preserving the existing project license. Official-source text retains its applicable source terms. Code on GitHub is MIT. Poster/third-party assets have separate attribution boundaries; they are linked, not relicensed as dataset content. See LICENSE-DATA.txt and the GitHub NOTICE.

Please credit Eugene Sun and Dupeng Liu, TrusOne-ComplianceKB, and identify the exact snapshot used. The original submission manuscript and the v5 poster describe different experimental stages. Conference presentation does not imply ACM proceedings publication; publication is not asserted.