heritage-ko-bench

Korean cultural-heritage knowledge benchmark v1.3 — 6 four-choice MCQ subsets (1,317 items) + 1 short-answer subset (270 items) = 1,587 items (test 1,552 · dev 35 for few-shot).

⚠️ Evaluation only. Do not train on this data.

Why this benchmark

• Decontamination by design: items are minted exclusively from 474 documents held out before the companion training suites were generated — document-level separation, empirically verified.
• Triple full-pass review: grounding verification → distractor back-verification → choices-only shortcut removal → answer-position balancing (per-subset A/B/C/D max−min ≤ 1) → three full manual-style audits → human review.
• Measured integrity: a 27B-class open model scores 61.2% zero-shot (subsets 57–67%), while choices-only accuracy is 29.8% ≈ random 25% (shortcut delta +31.4p) — items cannot be solved without reading the question.
• MMLU-style schema: choices list + 0-indexed integer answer — natively compatible with lm-evaluation-harness.

Subsets

subset	test	dev	topic
architecture	258	5	traditional architecture, palaces, temples
conservation	233	5	conservation science, restoration
excavation	240	5	archaeological excavation, artifacts
history_context	212	5	historical context, documents
intangible_folk	194	5	intangible heritage, folklore
natural_heritage	150	5	natural monuments, scenic sites
short_answer	265	5	all topics, short verifiable answers

Data Fields

MCQ subsets:

• question (string) — self-contained Korean question
• choices (list[string], 4) — options
• answer (int, 0-indexed: 0=A … 3=D)
• category (string) — subset name
• source_doc_id (string) — internal unit reference

short_answer: no choices; answer (string) is a short verifiable answer scored by exact match (case/punctuation-insensitive).

Running the benchmark (lm-evaluation-harness)

The lm_eval/ directory ships drop-in task configs (group heritage_ko + 7 tasks):

pip install "lm_eval[vllm]==0.4.8"   # pinned; newer versions have an --include_path regression (issue #3435)

# HF backend
lm_eval --model hf --model_args pretrained=<model>,dtype=bfloat16 \
  --tasks heritage_ko --include_path ./lm_eval \
  --num_fewshot 0 --batch_size auto --output_path results/ --log_samples

# vLLM backend (recommended)
lm_eval --model vllm --model_args pretrained=<model>,tensor_parallel_size=1,gpu_memory_utilization=0.8 \
  --tasks heritage_ko --include_path ./lm_eval --num_fewshot 0 --output_path results/

# Against a running OpenAI-compatible server (chat models)
OPENAI_API_KEY=EMPTY lm_eval --model local-chat-completions \
  --model_args model=<served-name>,base_url=http://<host>:8000/v1/chat/completions,num_concurrent=32 \
  --tasks heritage_ko --include_path ./lm_eval --apply_chat_template

Scoring: the 6 MCQ tasks use log-likelihood multiple_choice (acc/acc_norm) — suitable for base-model comparison. short_answer uses generate_until + exact_match (greedy, stop at newline). Run a single subset with e.g. --tasks heritage_ko_architecture.

Reproducibility anchors (fix these when comparing scores):

item	value
few-shot	report 0-shot by default; for 5-shot use fewshot_split: dev and say so
aggregation	group heritage_ko acc is size-weighted (micro) over the 6 MCQ tasks; report short_answer (exact_match) separately
generation scoring	short_answer: greedy (temperature 0), until: ["\n"]
harness version	report the lm-eval version alongside scores

How it was built

Held-out source units → teacher-LLM minting of MCQ/short items (source-grounded) → grounding verification, distractor back-verification, shortcut testing, dedup, convention cleanup, answer-position rebalancing → three full-pass item surgeries + human review.

License & Attribution

License: CC BY 4.0 — free to share and adapt with attribution.

Source material: The benchmark items were minted from Korean public-institution cultural-heritage reports published by the Korea Heritage Service (국가유산청) and the National Research Institute of Cultural Heritage (국립문화재연구원), distributed under the Korea Open Government License (KOGL, 공공누리).

KOGL type verification is complete for the Korea Heritage Service portion and still in progress for the NRICH-derived portion. Items are derived works (newly written questions and answer options), not redistributions of the source documents. If you are a rights holder and believe any item requires correction or removal, please open a discussion on this dataset page.

Attribution example

heritage-ko-bench (PoSTMEDIA AI Lab, 2026), CC BY 4.0.
Derived from Korean public-institution cultural-heritage reports (Korea Heritage Service /
National Research Institute of Cultural Heritage), Korea Open Government License.

Citation

@misc{heritage_ko_bench_2026,
  title     = {heritage-ko-bench: A Korean Cultural-Heritage Knowledge Benchmark},
  author    = {{PoSTMEDIA AI Lab}},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/PoSTMEDIA/heritage-ko-bench}
}