NOLLI

A reasoning benchmark of 15 puzzle categories in Korean and English, with three difficulty tiers calibrated so a reference model scores roughly 75% / 50% / 25% on easy / medium / hard.

• 25 subsets (category x language), 300 items each — 7,500 items total
• Each subset is one test split holding all three tiers (100 per tier)
• Every puzzle is programmatically generated with a verified unique solution

Usage

from datasets import load_dataset

ds = load_dataset("JoonYong-Park/NOLLI", "sudoku_ko", split="test")   # 300 items
print(ds[0]["question"])

# one difficulty tier
hard = ds.filter(lambda x: x["difficulty"] == "hard")     # 100 items

Rows are ordered easy (0-99), medium (100-199), hard (200-299), matching the numeric suffix of id.

Fields

field	type	description
id	string	{category}_{lang}_{0000-0299}
question	string	prompt shown to the model
answer	string	gold answer (always a string — see note)
solution	string	step-by-step reference derivation (STEP 0-STEP 3)
difficulty	string	easy / medium / hard
choices	string	multiple-choice options — kinship_ko only

answer is always a string. Numeric answers keep their exact form including leading zeros (e.g. number_baseball gold "0716432"). sat_puzzles stores its variable assignment as a JSON string — parse it with json.loads().

Subsets

subset	group	items
array_formula_en	Direct-translation	300
array_formula_ko	Direct-translation	300
causal_dag_en	Direct-translation	300
causal_dag_ko	Direct-translation	300
cipher_en	Script-specific	300
cipher_ko	Script-specific	300
cryptarithmetic_en	Script-specific	300
cryptarithmetic_ko	Script-specific	300
inequality_en	Direct-translation	300
inequality_ko	Direct-translation	300
jamo_ko	Korea-specific	300
kinship_ko	Korea-specific	300
korean_units_ko	Korea-specific	300
minesweeper_en	Direct-translation	300
minesweeper_ko	Direct-translation	300
number_baseball_en	Direct-translation	300
number_baseball_ko	Direct-translation	300
saju_ko	Korea-specific	300
sat_puzzles_en	Direct-translation	300
sat_puzzles_ko	Direct-translation	300
sudoku_en	Direct-translation	300
sudoku_ko	Direct-translation	300
time_ko	Korea-specific	300
yacht_dice_en	Direct-translation	300
yacht_dice_ko	Direct-translation	300

Evaluation

The model receives a system message describing the task and output format, and the question field as the user message. The final line is expected to be:

Answer: <your answer>

The reference grader also accepts 정답:, 최종 답:, and Final answer: as the label, and uses the last such line in the response.

Difficulty calibration

Difficulty is defined operationally: generator parameters were tuned until the reference model landed in the target bands (easy 65-85%, medium 40-60%, hard 15-35%, with at least 10%p between adjacent tiers). The knob differs per category — search space, constraint density, noise, or reasoning depth — so tiers are comparable in measured difficulty rather than in any single structural parameter.