KoEVD

KoEVD is a Korean benchmark linking five evaluation or analysis targets through source utterances: utterance-risk judgment, candidate-response safety choice, direct-generation response harmfulness, descriptive response strategies, and a pre-execution mock tool/action-choice diagnostic.

Contents and scope

The canonical corpus contains 13,552 sources and 71,395 response candidates: 30,740 accepted, 27,104 rejected, and 13,551 strongly rejected. Three interpretive fields record problem reasons, implied meanings, and background knowledge. Korean slang, idioms, relationships, and sociocultural references motivate depth within one language; this is not evidence of cross-lingual generalization.

All original JSONL files are byte-identical to the review package. See PACKAGE_README.md for paths and counts, schema/LABELS.md for labels, and METRICS.md for denominators. The repository manifest covers the current package; review_manifest.json preserves the earlier manifest for provenance.

Loading

Load release v0.1.0 using the explicit release tag:

from datasets import load_dataset
data = load_dataset("KETI-NLP/KoEVD", "canonical", revision="v0.1.0", split="test")

The sources_* configurations are source-ID indexes to join with canonical; they are not standalone utterance corpora. The core, hard, and balanced views overlap with the full corpus. They are not train/validation/test partitions. The test name only exposes evaluation records through the library API. Task 2 and Task 5 each use independent configurations because their schemas and variants differ. Task 2 relative_harmfulness has 13,551 rows; the other three variants each have 13,552.

Task 5 contains 784 selected sources under three framings (784 rows each). Its mock schemas are not executed. It is a diagnostic subset, not a population prevalence estimate. Task 4 reuses Task 3 generations and is descriptive, not a safety ranking.

Construction and governance

KoEVD is human-curated and AI-assisted, with guideline design, candidate drafting, rewriting, relabeling, adjudication and filtering. The final paper describes an approximately 3.5-year process and an 800-item IAA audit. No formal IRB approval or institutional exemption was obtained; the paper reports governance and annotator-welfare limitations explicitly. Labels reflect the project guidelines rather than universal moral ground truth.

Availability and limitations

The evaluation/analysis code, assistant-safety adapter, and response-strategy adapter are distributed separately under KETI-NLP. Release v0.1.0 identifies this dataset; the code release records immutable HF commits for GPU verification. B200 operational validation of the classifiers and offline CLI passed on fixed public samples; see the measured scope. This does not reproduce held-out F1. The strategy code/display-label crosswalk documents the original exporter mappings. Raw model-output dumps, internal construction documents, and the broader internal classifier-training pool are not part of this dataset. Classifier inference and exact classifier retraining therefore have different reproducibility scopes.

Use terms

The corpus intentionally includes harmful, offensive and sensitive language for safety research. Read LICENSE.md before use. The existing KoEVD research-use terms are preserved; this is not an unrestricted Apache/MIT dataset. Public examples should follow the sanitization requirements in those terms.

Citation

@misc{kim2026koevd,
  title={KoEVD: A Korean Benchmark for Metadata-Conditioned Multi-Stage LLM Safety Evaluation},
  author={Kim, San and Kim, Wonjung and Kwon, Heesoo and Yang, Jae Hee},
  year={2026},
  note={Accepted to the main conference of AACL-IJCNLP 2026; proceedings identifier pending}
}

Diagnose your own model

Use the complete KoEVD own-model workflow to generate your model’s responses, measure Task 1–5 behavior, and produce HTML/JSON/CSV reports of metadata-specific strengths, weaknesses, strategies and source-matched tool-selection gaps. The current GitHub source adds this pipeline; the original v0.1.0 code tag does not include it. A subsequent live B200 pipeline test completed 40 smoke requests and 959 expanded generation requests with unmodified code. It also found batch-dependent strategy outputs and concerning safety judgments; execution success does not establish measurement accuracy.

Historical Task 1 metadata view

The canonical corpus is unchanged. For exact historical Task 1 grouping, the accompanying metadata overlay records 35 sources whose severity or utterance type differs from the retained analysis view. The checksum-checked script emits a separate metadata index; it does not overwrite canonical labels or model inputs. The index has 13,430 non-borderline sources, compared with 13,429 in the canonical release. Record the annotation view when comparing results.

A later B200 check reproduced the assistant-safety classifier on all 532 original validation inputs; it did not establish accuracy on new generations or reproduce strategy validation F1. See the updated reproduction scope.