datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
GridCorpus_9M_Sudoku_Puzzles_Enriched
╔══════════════════════════════════════════════════════════════════════╗
║ ║
║ G R I D C O R P U S ║
║ ║
║ "004300209005009001070060043..." ║
║ │ ║
║ ▼… See the full description on the dataset page: https://huggingface.co/datasets/beta3/GridCorpus_9M_Sudoku_Puzzles_Enriched.puzzlescript-gists
PuzzleScript Human-Authored Games (Full Gist Corpus)
35,704 human-authored PuzzleScript games — the
complete source text of each — collected from public GitHub gists.
This is the full corpus: every distinct gist is kept, and each row is tagged
with its deduplication cluster so you can reduce to a unique set with a one-line
filter. The deduplication is reproducible from the shipped dedup_master.json +
dedup_master.py; non-vanilla PuzzleScript-Plus files are excluded (listed in… See the full description on the dataset page: https://huggingface.co/datasets/smearle/puzzlescript-gists.chess_puzzle_benchmark
Chess Puzzle Benchmark
Chess puzzle evaluation sets at five difficulty tiers, B1 (easiest) through
B5 (hardest). Each example is a chess game given in PGN move notation; the
model must produce the next move(s).
Two prompt variants
The same puzzles are released in two forms that differ only in the prompt suffix:
think/ — the prompt ends with a special <T> token. <T> is a
reasoning trigger: it tells the model to think (produce a chain of reasoning)
before… See the full description on the dataset page: https://huggingface.co/datasets/pavelslab-nyu/chess_puzzle_benchmark.GLM-5.2-Logic-Puzzles
GLM-5.2 · Logical Puzzles
6000x traces distilled from GLM-5.2 on High reasoning
Token Count: 5M~?
Distribution:
Puzzles:
•Tokenization blindless ex: counting the r's in strawberry
•Goal reasoning ex: the car wash test (theres no car wash question exactly just prompts like it so its not just benchmaxxing)
•Reading comprehension traps
•Temporal reasoning
•Many other categories not worth mentioning
Prompts… See the full description on the dataset page: https://huggingface.co/datasets/ianncity/GLM-5.2-Logic-Puzzles.logic-grid-puzzles-training-pool
Logic grid puzzles training pool
Logic grid puzzles: a row of positions, a handful of attributes with one value per position, and a
list of clues that together admit exactly one arrangement. Two sets drawn for this pool by
generators run here under the seeds recorded below, and two public datasets read at the pinned
revisions named below, laid out twice. Train on either layer or on both.
pool.jsonl
Every source rewritten into one shape, 390945 rows, one JSON… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/logic-grid-puzzles-training-pool.zebra-puzzlesSynthetic data for the paper [2505.05755] Insertion Language Models: Sequence Generation with Arbitrary-Position Insertions.
Project page: https://dhruveshp.com/projects/ilm
puzzle-bench
Puzzle Bench
Difficulty-labeled evaluation datasets for Sudoku and Maze tasks, designed for benchmarking language models on combinatorial reasoning.
GitHub: zeyuzhangzyz/puzzle-bench
Dataset Overview
Config
Total
Train
Test
Difficulty Labels
sudoku_4x4
10,000
8,000
2,000
--
sudoku_9x9
52,806
42,244
10,562
easy / medium / hard
maze_5x5
10,000
8,000
2,000
--
maze_7x7
10,000
8,000
2,000
--
maze_10x10
10,000
8,000
2,000
--
maze_15x15
30,000
24,000
6,000… See the full description on the dataset page: https://huggingface.co/datasets/zeyuzy/puzzle-bench.combi-puzzles
Combi-Puzzles Dataset
This repository contains the Combi-Puzzles dataset used in the research paper titled "Can Language Models Rival Mathematics Students? Evaluating Mathematical Reasoning through Textual Manipulation and Human Experiments." These variations are designed to evaluate problem-solving strategies across different formats.
Dataset Description
The Combi-Puzzles dataset includes 125 problems:
25 Base Combinatorial Problems: Covers permutations… See the full description on the dataset page: https://huggingface.co/datasets/andynik/combi-puzzles.rukh-puzzles-split
chorcat/rukh-puzzles-split
Lichess puzzles with rating deviation <= 100 and at least 100 plays, banded by difficulty (1000-1500, 1500-2000, 2000+) and split into test and train by a seeded hash of the puzzle id, each with the moves of the game it came from, for tactical evaluation and fine-tuning.
Part of Rukh, a chess language model built from scratch
as a course on generative and agentic AI. Every derived dataset ships with the exact filters and
counts of its manifest.json, so… See the full description on the dataset page: https://huggingface.co/datasets/chorcat/rukh-puzzles-split.Einstein-Puzzles-Data
Einstein-Puzzles
Communication and Verification in LLM Agents towards Collaboration under Information Asymmetry (Arxiv)
Run Peng*, Ziqiao Ma*, Amy Pang, Sikai Li, Zhang Xi-Jia, Yingzhuo Yu, Cristian-Paul Bara, Joyce Chai
Dataset Details
There are four *.jsonl files under train/ folder, which corresponds to training data for models with four different communicative action spaces. The chain-of-thought reasoning traces are generated by gpt4o given the current game state and… See the full description on the dataset page: https://huggingface.co/datasets/Roihn/Einstein-Puzzles-Data.knights-knaves-puzzles
Knights and Knaves Logic Puzzles Dataset
A comprehensive dataset of Knights and Knaves logic puzzles ranging from 3 to 14 inhabitants.
Each puzzle requires logical deduction to determine who tells the truth (knights) and who lies (knaves).
The dataset includes detailed chain-of-thought reasoning for each solution.
Dataset Description
This dataset contains 12,000 Knights and Knaves logic puzzles. In these puzzles:
Knights always tell the truth
Knaves always lie
The goal… See the full description on the dataset page: https://huggingface.co/datasets/RedaAlami/knights-knaves-puzzles.lichess_puzzle_datasetLichess Puzzle Database
Updated for March 2025.
chess_puzzle_training_datasets_lt-2400
Chess puzzle training datasets: rating below 2400
This is a filtered derivative of
pavelslab-nyu/chess_puzzle_training_datasets.
Every retained row satisfies the exact condition:
Rating < 2400
Rating is the Lichess puzzle rating, not the Elo of either player in the
source game. The original column names, column order, directory layout, and CSV
schemas are preserved. As in the upstream repository, Hugging Face discovers
all three CSVs as one default configuration with one train… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/chess_puzzle_training_datasets_lt-2400.chess-debate-puzzles
Chess Debate Puzzles
A stratified sample of Lichess mid/endgame chess puzzles annotated with Stockfish-evaluated
moves across ten centipawn-quality bands. Designed for experiments in the spirit of
AI Safety via Debate (Irving et al., 2018), where two AI
agents argue for different moves and a judge must identify the objectively better one.
Motivation
Debate as an alignment technique asks whether a human (or AI) judge can identify the correct
answer when two agents argue… See the full description on the dataset page: https://huggingface.co/datasets/kvoudouris/chess-debate-puzzles.mini-swe-rl-puzzles
Mini-SWE-RL — puzzles, rollouts, and replication artifacts
Everything needed to compare a mini-swe-rl-grpo run against ours,
and to diagnose whether a result difference is training randomness or a setup difference.
Trained checkpoint: sriharsha4444/mini-swe-rl-grpo-qwen2.5-coder-1.5b
Contents
File
What it is
puzzles.jsonl
All 45 puzzles (id, tier, difficulty, description, buggy_code, tests) with flags for the 16-puzzle GRPO train set and the 30-puzzle… See the full description on the dataset page: https://huggingface.co/datasets/sriharsha4444/mini-swe-rl-puzzles.puzzlezoo
PuzzleZoo
PuzzleZoo is a collection of three reasoning / planning benchmarks designed to evaluate large language models on multi-step procedural problem-solving — the kind of task where one wrong primitive move silently invalidates the rest of the plan.
It is the official evaluation suite for the paper RePoT: Recoverable Program-of-Thought via Checkpoint Repair (Mazaheri, 2026, arXiv:2605.30052), and is released as a standalone benchmark for the broader community.… See the full description on the dataset page: https://huggingface.co/datasets/parsa-mz/puzzlezoo.15-puzzle
15 Puzzle
This repository contains SFT, RL, and evaluation data for the 15-puzzle task.
Files
data/sft-00000-of-00001.parquet: 600 supervised fine-tuning puzzles
data/eval-00000-of-00001.parquet: 100 held-out evaluation puzzles
data/rl-00000-of-00001.parquet: 10,000 reinforcement learning puzzles
SFT data
The SFT split contains 600 puzzles across scramble depths 8 through 15, with
75 puzzles at each depth. Each puzzle was solved with an optimal… See the full description on the dataset page: https://huggingface.co/datasets/saad1926q/15-puzzle.chess-puzzle-cot
Chess Puzzle Chain-of-Thought (augmented)
Chain-of-thought traces for one-move Lichess puzzles (beginner bucket, rating
399–999). This is the augmented CoT set used as SFT data for the chess
reasoning models, oversampling weak tactical families (hanging pieces, pins,
advanced pawns, promotions, discovered/double checks, and scarce
advantage/crushing material wins).
Schema
field
type
description
puzzle_id
string
Lichess puzzle id
rating
int
Lichess… See the full description on the dataset page: https://huggingface.co/datasets/gshikhar/chess-puzzle-cot.