NOLLI

NOLLI (논리) is a benchmark of rule-based logical puzzles in Korean and English, designed with difficulty calibration in mind. It covers 15 task types x 3 difficulty tiers (easy / medium / hard) x 100 examples each, for a total of 7,500 examples (EN 3,000 + KO 4,500).

All examples are programmatically generated with solver-verified answers. If you need more data, the generators in the GitHub repository can produce additional examples for any task, language, and difficulty tier.

📄 Paper: https://arxiv.org/abs/2608.04397
💻 Code: https://github.com/HAE-RAE/NOLLI

Tasks

Tasks follow a three-level cross-lingual spectrum: direct translations, script adaptations, and Korean-only tasks.

1. Direct Translations (8 tasks · EN + KO)

Same generators and parameters in both languages, so performance gaps isolate presentation language.

task	description	en (easy/med/hard)	ko (easy/med/hard)	total
array_formula	Apply row/column aggregation formulas (SUM/MEAN/MAX/...) to a 2D array in sequence	100 / 100 / 100	100 / 100 / 100	600
causal_dag	Infer event-propagation time through a causal graph with per-edge delays	100 / 100 / 100	100 / 100 / 100	600
inequality	CSP: place numbers 1..N satisfying inequality constraints (unique solution)	100 / 100 / 100	100 / 100 / 100	600
minesweeper	Minesweeper with minimal hints preserving a unique solution	100 / 100 / 100	100 / 100 / 100	600
number_baseball	Infer a hidden N-digit number from Strike/Ball hints	100 / 100 / 100	100 / 100 / 100	600
sat_puzzles	Boolean satisfiability (CNF) framed as a natural-language scenario	100 / 100 / 100	100 / 100 / 100	600
sudoku	9x9 Sudoku with guaranteed-unique solutions	100 / 100 / 100	100 / 100 / 100	600
yacht_dice	Assign 12 dice rolls to 12 scoring categories to maximize total score; report the partial sum over spot-checked rounds	100 / 100 / 100	100 / 100 / 100	600

2. Script Adaptations (2 tasks · EN + KO)

Not translation-equivalent: English uses Roman letters; Korean uses Hangul jamo. Calibrated independently per language.

task	description	en (easy/med/hard)	ko (easy/med/hard)	total
cipher	Decode stacked ciphers (Substitution, Vigenere, Reverse, Playfair, Transposition)	100 / 100 / 100	100 / 100 / 100	600
cryptarithmetic	Letters-for-digits arithmetic (SEND+MORE=MONEY style); Korean maps jamo groups	100 / 100 / 100	100 / 100 / 100	600

3. Korean-Only (5 tasks · KO)

No English counterpart — built on structures specific to the Korean language and culture (kinship terminology, the sexagenary calendar, traditional units, Hangul orthography).

task	description	ko (easy/med/hard)	total
jamo	Decompose a Hangul syllable into 초성/중성/종성, shift 초성, recompose	100 / 100 / 100	300
kinship	Infer a Korean kinship term from a chain of family relationships (26-way multiple choice)	100 / 100 / 100	300
korean_units	Convert traditional Korean units (평·마지기·되·자·냥, ...) with randomized rates	100 / 100 / 100	300
saju	Compute the four 사주(四柱) pillars from a birth date/time	100 / 100 / 100	300
time	Korean calendar reasoning: holiday anchor + relative-day expression → date / 60갑자 일진	100 / 100 / 100	300

Usage

from datasets import load_dataset

# one task
ds = load_dataset("HAERAE-HUB/NOLLI", "sudoku", split="test")

# filter by language / difficulty (both are columns)
ko_hard = ds.filter(lambda x: x["language"] == "ko" and x["difficulty"] == "hard")

Structure

Each config corresponds to one task. All configs share a single test split and the same schema (the choices column exists only in kinship):

column	type	description
id	str	unique example id, e.g. sudoku_ko_0000
task	str	task name (same as config name)
language	str	ko or en
difficulty	str	easy / medium / hard
question	str	full prompt for the model
choices	str	JSON-encoded answer choices — only present in the kinship config
answer	str	gold answer
solution	str	step-by-step reference solution

Citation

@misc{choi2026nolli,
      title={NOLLI: A Difficulty-Calibrated Puzzle Benchmark for Diagnosing the English-Korean Performance Gap},
      author={Dasol Choi and Joonyong Park and Daegon Yu and Soo Yong Kim and Youngsook Song and Seunghyeok Hong},
      year={2026},
      eprint={2608.04397},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.04397},
}