max-rl/knk_n3_12

Knights and Knaves, n_people = 3..12

Procedurally generated Knights-and-Knaves logic puzzles for RL with a binary reward, in verl parquet format. Every puzzle has a unique solution.

file	split	rows	n_people
train.parquet	train	100,000	3..12, 10,000 each
eval.parquet	validation	1,000	3..12, 100 each, stratified over (depth, width) cells
fahim_test_n10.parquet	fahim_test_n10	1,000	10 (the test split of ftajwar/knights_and_knaves_fraction_reward, same prompt format)

How it was made

Same procedure as ftajwar/knights_and_knaves_fraction_reward (reverse-engineered from its rows), scaled from 20k to 100k and to n_people 3..12:

• Generator: reasoning_gym 0.1.19 KnightsKnavesDataset (the K&K sampler of Xie et al., 2024). Item idx of a cell is drawn from Random(cell_seed + idx), and each cell has its own seed.
• (depth, width) grid: for n ≤ 5, {(2,3), (3,2), (3,3)}. For n ≥ 6, {2,3,4}×{2,3,4} plus (5,4), which gives 10 cells. Train and eval are split evenly across each n's cells.
• Dedup: within each cell, idx runs 0, 1, 2, …, and any puzzle whose statements equal an earlier one in the same n is skipped. n = 2 is not included because one of its cells has only about 975 distinct puzzles in total.
• No overlap: no duplicate statements within train or eval. No train puzzle is isomorphic to an eval puzzle or to a fahim_test_n10 puzzle, where isomorphic means equal after relabelling the people and reordering the operands of and / or / <=>. Train/eval prompt-text overlap is 0.

Reproduce with generate_knights_and_knaves_data.py (included here, see its docstring for details):

pip install reasoning-gym==0.1.19 pandas pyarrow
python generate_knights_and_knaves_data.py --out_dir OUT --fahim_dir DIR_WITH_FAHIM_{train,test}.parquet \
    --n_people 3,4,5,6,7,8,9,10,11,12 --train_per_n 10000 --eval_per_n 100 --auto_shift \
    --seed 20261007 --prompt_style math

stats.json records every cell's seed, number of draws, distinct-puzzle count, train/eval counts, and the overlap checks.

Row format

{
  "data_source": "knights_and_knaves"                 # train
               | "knights_and_knaves_eval_nXX"        # eval, one per n (per-difficulty metrics in verl)
               | "knights_and_knaves_fahimtest_n10",
  "prompt": [{"role": "user", "content": question + "\nPlease reason step by step, and put your final answer within \\boxed{{}}."}],
  "ability": "logic",
  "reward_model": {"style": "rule", "ground_truth": "<JSON: statements, solution, names, knight_knave_terms, difficulty, answer, question, ...>"},
  "extra_info": {"split", "index", "n_people", "depth_constraint", "width_constraint", "cell_seed", "source_index"},
}

• Prompt: the question already contains the answer format, e.g. (Format your answer like: "A is a hero/villain, ..."). The appended instruction is copied verbatim from the maxrl math preprocessing, including its literal \boxed{{}}, so KnK and math runs share one prompt.
• Reward (binary): take the last \boxed{...} (one extra wrapping {...} layer is tolerated), parse it into a set of (name, role) pairs with reasoning_gym's _normalize_answer, and return 1.0 iff that set equals the set parsed from ground_truth["answer"], else 0.0.