KR-CompBench

Version 1.1.0 · 624 English scenarios · 312 pairs · 14 fields

KR-CompBench is a benchmark for binary compliance assessment against Korean Register (KR) classification rules, guidance, and interpretations. Each case combines a technical scenario and assessment scope with a gold label, source-linked checks, and available normative dependency annotations. It supports research on maritime regulatory reasoning, retrieval-assisted assessment, and decisions that depend on linked provisions.

This release updates the dataset previously named KRcompliance. Its Hugging Face repository remains jiseongch/KRcompliance for compatibility.

Loading

import json
from datasets import load_dataset

dataset = load_dataset("jiseongch/KRcompliance", revision="v1.1.0")
train, dev, test = dataset["train"], dataset["dev"], dataset["test"]
assert (len(train), len(dev), len(test)) == (436, 94, 94)

# Load an individual partition when needed.
test = load_dataset("jiseongch/KRcompliance", split="test", revision="v1.1.0")
example = test[0]
model_input = {key: example[key] for key in ("task_scope", "scenario")}

# Gold annotations belong with the evaluator.
gold_label = example["label"]
gold_checks = json.loads(example["checks_json"])

JSON columns contain JSON-encoded strings. Decode them explicitly. Empty scalar cells may load as null; when reading CSV directly, use encoding="utf-8-sig".

Task and labels

Given task_scope and scenario, predict exactly compliant or non-compliant under the applicable source edition. A non-compliant case has at least one check marked not_met; a compliant case has none. A check verdict of not_applicable is not a third case label. Labels concern the stated assessment scope, not whole-ship approval.

All other columns are evaluator-side metadata. Exclude id, case and pair identifiers, pair types, labels, quoted evidence, check assessments, and gold graphs from ordinary prediction prompts. For retrieval experiments, state the source corpus and retrieval protocol. Supplying case-specific gold evidence constitutes an oracle-evidence setting and should be reported separately.

pair_type describes the two final gold labels: contrast means they differ; control means they agree. It does not encode original construction intent. There are 300 contrast pairs and 12 control pairs, including nine compliant/compliant and three non-compliant/non-compliant pairs.

Splits and coverage

Measure	Train	Dev	Test	Total
Cases	436	94	94	624
Pairs	218	47	47	312
Compliant cases	222	48	48	318
Non-compliant cases	214	46	46	306
Contrast pairs	214	42	44	300
Control pairs	4	5	3	12
1-hop cases	234	51	51	336
2-hop cases	80	17	17	114
3-hop cases	20	4	5	29
4-hop cases	2	0	1	3
5-hop cases	2	0	0	2
Hop unannotated	98	22	20	140

All 11 broad task categories occur in each split. They cover classification surveys; materials, welding and non-destructive testing; hull structures and outfitting; machinery, piping and pressure vessels; electrical installations and control; batteries; liquefied-gas carriers; low-flashpoint fuels; fire safety; additional installations; and pollution prevention. These strata differ from the 302 original task_scope values and 252 rule_family codes.

Pairs and 151 connected holdout groups remain intact. Recorded checks_json.unit_id values and rule-family codes are disjoint across splits. Exact identifiers and mechanically normalized aliases have no cross-split overlap. Source documents can be shared; these checks do not rule out broader semantic equivalence, shared parent requirements, or pretraining exposure.

All 624 cases were already published in the preceding dev-only release; the new partitions define recorded-clause and rule-family holdouts without establishing that test cases were previously unseen or unexposed.

The 484 annotated cases include 336 one-hop and 148 multi-hop cases; 39 require cross-document reasoning. The remaining 140 cases are unannotated. The only 5-hop pair is in train, and dev has no 4-hop cases. The release therefore does not support a separate 5-hop test-performance estimate.

Category definitions and full distributions are in dataset_metadata.json. Per-case task categories and holdout groups are in metadata/case_metadata.jsonl.

Fields

Field	Description
id	Integer identifier from 1 to 624, assigned in the original master CSV row order and retained across splits.
case_id	Unique case identifier.
pair_id	Identifier shared by two related cases.
pair_type	Final gold-label relationship: contrast or control.
split	train, dev, or test.
task_scope	Scope of the compliance assessment; model input.
scenario	Technical case description; model input.
label	compliant or non-compliant.
rule_family	Fine-grained rule-family code.
checks_json	Checks with requirements, assessments, verdicts, source identifiers and hashes, locators, PDF pages, quotations, and quotation provenance.
hop_count	Normative node count on the longest recorded directed dependency path, including endpoints; empty when unannotated.
requires_cross_document_reasoning	Recorded cross-document annotation; empty when unannotated.
reasoning_nodes_json	Normative dependency nodes and source references.
reasoning_edges_json	Directed relations between dependency nodes.

id is the first CSV column. Its mapping to case_id is global and consistent between the master CSV and partitions; numbering does not restart within each split. Existing case_id values are preserved and remain the identifiers used by the evaluator.

Hop count measures path depth in nodes. It is not the number of checks, total graph nodes, or edges; independent branches do not add to depth. Missing hop and cross-document values must remain missing rather than becoming zero, one-hop, or false. Unannotated graphs are empty lists, [].

Sources and provenance

The 1,082 checks reference 279 distinct recorded source-unit identifiers and 18 source-document identifiers. Sixteen documents are identified as 2025 editions, with a 2022 interpretation casebook and 2023 remote-survey guidance also represented. Coverage includes KR rules and guidance for hulls, machinery, electrical systems, specialized ships, and additional installations.

There are 1,080 quotations marked verbatim English and two marked English translations of official Korean text. The release includes source excerpts in checks_json[].quote. Source identifiers, hashes, clause locators, and physical PDF page numbers are retained in the checks; PDF pages are not printed page labels. Recorded provenance does not establish that an edition remains current or that every quotation has been independently reverified.

Construction procedure

The following procedure is documented from the dataset creator's account. Model names identify the reported generation and review tools.

1. Build the policy graph. Convert KR documents into a graph of normative provisions and their relationships.
2. Establish sources and seed scenarios. Collect KR rules, guidance, and interpretation casebooks; verify provenance and applicable editions; obtain regulatory-application scenarios for compliant cases.
3. Construct reference chains. Include those scenarios and build the relevant chains of referenced provisions from the policy graph.
4. Expand task coverage. Create additional scenario chains for tasks not yet represented.
5. Generate labeled data points. Use Claude Opus 5.5 to convert scenarios into cases with compliant and non-compliant applications.
6. Make meaningful multi-hop extensions. Connect referenced clauses, parent provisions, definitions, exceptions, and additional requirements to create situations in which those dependencies change the judgment.
7. Review scenarios and dependencies. Use a separate GPT-6.1 review agent to reassess scenarios, labels, applicability, and reference relationships, with review isolation intended to reduce contamination.
8. Review answer-revealing cues. Manually inspect and revise overt wording or excessive information that could reveal the label without consulting the relevant KR provisions.
9. Expert evaluation — Ongoing. Expert assessment is in progress; final results and agreement metrics are not yet available.

The v1.1.0 procedure has no separate dummy/noise-sentence insertion stage. Reviewer separation does not establish label-blinded adjudication or freedom from contamination. Earlier documentation records AI review with access to existing labels. Expert evaluation remains ongoing.

Evaluation

The included evaluator reports case accuracy, macro-F1, balanced accuracy, per-label precision/recall/F1, and a confusion matrix. Pair metrics are:

• Pair joint accuracy: both predictions are correct, over all pairs and separately by pair type.
• Contrast flip consistency: predicted labels differ, over contrast pairs only.
• Control prediction stability: predicted labels agree, over control pairs only.

Consistency alone does not establish correctness. Report it alongside case and pair joint accuracy. The test denominators are 94 cases, 47 total pairs, 44 contrast pairs, and three control pairs; the small control subset limits conclusions.

Create predictions.csv with columns case_id,prediction, using one row per selected case and the exact two label strings. Download the repository, then run:

git clone --branch v1.1.0 https://huggingface.co/datasets/jiseongch/KRcompliance
cd KRcompliance
python evaluate_predictions.py --predictions /path/to/predictions.csv --split test --output metrics.json

The script uses the Python standard library and loads the three data/*.csv files by default. Supply an alternative master CSV with --data. It requires every selected case's prediction and validates pair types against gold labels.

Report the version, split, input fields, retrieval setting, and averaging method. Keep unannotated cases separate in hop-based analyses. Explanation quality requires a separately defined assessment; no validated automatic explanation scorer or headline model results are supplied.

Release files and changes

• data/train.csv, data/dev.csv, data/test.csv: canonical partitions, containing each case once.
• dataset_metadata.json and metadata/case_metadata.jsonl: aggregate and per-case metadata.
• evaluate_predictions.py: case and pair evaluation.
• KR2025_v1.1.0.html: downloadable standalone case/evidence viewer with filtered CSV export.
• validation_summary.json and SHA256SUMS: mechanical validation and release-file checksums.

Compared with v1.0.12-en1, v1.1.0 introduces the KR-CompBench name, reduces the CSV schema from 33 to 14 fields, adds pair_type and the sequential numeric id, and replaces the all-dev layout with train/dev/test partitions. Splits change for 530 cases. The 624 scenario texts, labels, checks, recorded graphs, and existing case_id values are retained. The release adds task/holdout metadata and pair-aware evaluation, and updates the documented construction procedure to omit the noise-insertion stage.

Mechanical validation establishes file and schema integrity; it does not replace semantic or expert assessment.

Limitations and rights

Model-assisted generation and label-aware review may retain errors or confirmation bias. Coverage is limited to the represented editions and selected tasks. Label balance is a dataset property, not an estimate of real-world non-compliance prevalence. Synthetic scenarios, translations, and incomplete reasoning annotations can affect evaluation. This benchmark is not official KR certification or a substitute for expert assessment of an actual vessel.

No open license is specified for creator-authored dataset content. Third-party source quotations retain their respective rights holders' rights; this card grants no rights to those excerpts. No redistribution-permission statement for quoted KR material is included. Users should establish the applicable permissions for their intended reuse.

Citation

@misc{jiseongch2026krcompbench,
  title   = {KR-CompBench},
  author  = {jiseongch},
  year    = {2026},
  version = {1.1.0},
  url     = {https://huggingface.co/datasets/jiseongch/KRcompliance}
}