RegConflict

A cross-jurisdictional benchmark for regulatory conflict detection.

RegConflict is an expert-annotated benchmark for detecting conflicts between regulatory regimes from different jurisdictions in financial regulation. It pairs a four-class typology of conflict kinds with labelled regime pairs, each grounded in verbatim evidence passages drawn from a curated multi-jurisdictional regulatory corpus.

• Source code, corpus, and reproducible baselines: https://github.com/rafsunsheikh/RegConflict
• Archived release (DOI): 10.5281/zenodo.21007776

This HuggingFace repository is the dataset mirror — the labelled regime pairs with their embedded evidence passages. The full document corpus and baselines live in the GitHub repository and the Zenodo deposit.

The task

Given two regulatory regimes (each represented by evidence passages drawn from official source documents), predict whether they conflict or do not conflict, and — for conflicts — which kind of conflict applies under a four-class typology distinguished by the resolution mechanism (or its absence) available to an entity facing both regimes:

• structural_unresolved — no compliant path exists under the rules as written.
• operationally_resolvable — a compliant path exists but requires deliberate operational structuring.
• interpretive_fact_sensitive — whether a conflict exists turns on facts or interpretation.
• recurring_friction — an ongoing economic/operational friction with no clean one-time fix.

Splits

Split	Records	Conflict	Non-conflict
train	497	65	432
validation	107	14	93
test	107	14	93
gold_iaa	30	25	5

gold_iaa is a separately-held gold inter-annotator-agreement subset of conflict cases, blind-reviewed to measure agreement on the four-class typology task (Cohen's κ = 0.6 — substantial agreement per Landis–Koch). It carries an additional blind_review field.

Usage

from datasets import load_dataset

# Main benchmark: train / validation / test
ds = load_dataset("rafsunsheikh/RegConflict")
print(ds)
print(ds["test"][0]["label"])

# Gold inter-annotator-agreement subset (separate config, richer schema)
gold = load_dataset("rafsunsheikh/RegConflict", "gold_iaa")
print(gold["gold_iaa"][0]["blind_review"])

The gold_iaa subset is exposed as a separate config rather than a fourth split of default, because it carries an extra blind_review field whose schema differs from the main splits.

Record schema

Each record is one regime pair. Key fields (some are present only for one label):

Field	Type	Notes
pair_id	string	Stable identifier, prefixed conflict: / non_conflict:.
label	string	conflict or non_conflict.
record_type	string	Same as label.
regime_a, regime_b	struct	regime_id, jurisdiction, issuing_body, short_name.
evidence_a, evidence_b	list<struct>	Evidence passages: chunk_id, source_doc, page, passage.
conflict_type	string|null	One of the four classes (conflict records only).
severity	string|null	low / medium / high (conflict records only).
rationale	string	Free-text justification of the label.
compliant_paths	list<string>	Resolution options (conflict records).
out_of_scope_assumptions	list<string>	Scoping assumptions (conflict records).
non_conflict_subtype	string|null	e.g. different_domains, aligned_complementary (non-conflict records).
would_conflict_if	list<string>	Counterfactual conditions (non-conflict records).
confidence	float	Annotator confidence.
annotator_id, annotated_at, revision, review_status	—	Provenance.
blind_review	struct	Blind second-pass review (gold_iaa split only).

Jurisdictions and sources

Pairs are drawn from six jurisdictional sources — Australia, the European Union, Singapore, FATF, BIS, and IOSCO — spanning bodies such as APRA, OAIC, AUSTRAC, ASIC, the Australian Treasury, EUR-Lex (GDPR, MiCA, AMLA, AMLR/TFR, DORA, …), ESMA/EBA, the Singapore SSO and MAS, the Basel Committee, and IOSCO.

Licensing

• Labels, annotations, and this dataset structure: MIT (see the GitHub repository's LICENSE).
• Embedded evidence passages are short verbatim quotations from official regulatory documents, included under academic-quotation terms. The full source corpus uses a three-tier license scheme (Tier 1 redistributable, Tier 2 academic-use, Tier 3 metadata-only); see LICENSES/SOURCE_DOCUMENT_LICENSES.md in the GitHub repository for verbatim per-source clauses.

Citation

Please cite the dataset deposit; see CITATION.cff for the canonical entry.

@dataset{regconflict,
  title     = {RegConflict: A Cross-Jurisdictional Benchmark for Regulatory Conflict Detection},
  author    = {Sheikh, MD Rafsun},
  year      = {2027},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21007776},
  url       = {https://doi.org/10.5281/zenodo.21007776}
}