PuzzleClone Dataset

A comprehensive, diverse, and verifiable dataset of over 83k logical puzzles generated by PuzzleClone.

• 📜 Paper: https://arxiv.org/abs/2508.15180

🌟 Dataset Overview

PuzzleCloneData contains 83,657 unique logical reasoning puzzles procedurally generated from 86 seed puzzles. The dataset spans:

• Various applications of Satisfiability Modulo Theories (SMT) and SMT-like puzzles,
• Classic logical puzzles like Sudoku, the Knapsack problem, and linear optimization (LP).
• Diverse mathematical problems of varying difficulties.

Remarkable features:

• ✅ Guaranteed Verifiability: Every problem is generated with a ground-truth solution and is verifiable by a symbolic SMT solver, ensuring correctness.
• 🎯 Granular Control: Offers fine-grained control over problem attributes like scale, structure, and difficulty through a set of adjustable parameters, enabling large-scale batch generation.
• ✨ Flexible Adaptation: Facilitates the easy customization of problem scenarios and translation into different languages or domains.

🏗️ Dataset Structure

PuzzleClone/
├── normal/           # Normal difficulty problems
│   ├── RL_train      # Reinforcement Learning training data
│   ├── RL_validate   # Reinforcement Learning validation data
│   ├── SFT           # Supervised Fine-Tuning data
│   └── Test          # Test data
├── hard/             # Hard difficulty problems
│   ├── RL_train      # Reinforcement Learning training data
│   ├── RL_validate   # Reinforcement Learning validation data
│   ├── SFT           # Supervised Fine-Tuning data
│   └── Test          # Test data
└── readme.md

Usage

from datasets import load_dataset

# Load the normal difficulty subset
ds_normal = load_dataset("hithink-ai/PuzzleClone", "normal")

# Check available splits
print(ds_normal)
# Example output:
# DatasetDict({
#     RL_train: Dataset(...),
#     RL_validate: Dataset(...),
#     SFT: Dataset(...),
#     Test: Dataset(...)
# })

# Access a single example from the SFT split
sample = ds_normal["SFT"][0]

# Iterate over the RL_train split
for example in ds_normal["RL_train"]:
    problem_text = example["problem"]
    answer_text = example["answer"]
    # Can be used for training RL models or generators
    ...

# Load the hard difficulty subset
ds_hard = load_dataset("hithink-ai/PuzzleClone", "hard")
print(ds_hard["Test"][0])

# If you only want to keep specific fields
ds_sft = ds_normal["Test"].remove_columns(["config", "parameters"])
print(ds_sft[0])

📊 Key Statistics

General Statistics

Split	Normal Difficulty	Hard Difficulty	Total
RL Training	50,738	23,616	74,354
RL Validation	430	430	860
SFT Training	2,161	2,139	4,300
Test	5,730	2,713	8,443
Total (*)	56,898	26,759	83,657

*SFT training samples are selected from other training splits and thus are not included in the overall total.

Difficulty Distribution

Puzzle complexity is quantified using a composite difficulty score (between 0 and 1, higher means more difficult), calculated as the normalized average of four features: the number of random variables, the number of logical constraints, the problem description length, and a custom variable domain difficulty metric (var scale).

The difficulty distribution of the puzzles is visualized below.

Question Type Distribution

File (*)	简答题 (Short Answer)	单选题 (Multiple-Choice)	填空题 (Fill-in-the-Blank)
normal/RL_validate	390	95	10
normal/Test	5,108	1,262	124
normal/SFT	1,966	475	50
normal/RL_train	45,195	11,202	1,099
hard/RL_validate	390	95	10
hard/Test	2,514	644	78
hard/SFT	1,934	475	50
hard/RL_train	21,889	5,631	679
Total (**)	75,486	18,929	2,000

*If a puzzle has multiple sub-questions, the question types of these sub-questions will be separately counted.

**SFT training samples are selected from other training splits and thus are not included in the overall total.

Answer Type Distribution

PuzzleClone supports 8 different types of puzzle answers:

• numeral: Numeric answers (e.g., a number or a list of numbers).
• option: Single-letter answers (e.g., 'A', 'B') representing the selected option.
• ordered array: Answers that are arrays where the order of elements matters.
• nominal: Answers that are categorical labels or names.
• unordered array: Arrays where the order of elements does not matter.
• ooa_numeral: Two-dimensional arrays of numeric values where the orders of elements in both dimensions matter.
• ooa_nominal: Two-dimensional arrays of nominal values where the orders of elements in both dimensions matter.
• oua_nominal: Two-dimensional arrays of nominal values where only the order of elements in the outer dimension matters.

The following table shows the distribution of the answer types in PuzzleClone.

File (*)	numeral	option	ordered array	nominal	unordered array	ooa_numeral	ooa_nominal	oua_nominal
normal/RL_validate	180	95	80	70	55	5	5	5
normal/Test	2,089	1,262	1,232	961	727	55	77	91
normal/SFT	910	475	404	352	275	25	25	25
normal/RL_train	18,444	11,202	10,936	8,503	6,442	483	681	805
hard/RL_validate	180	95	80	70	55	5	5	5
hard/Test	1,501	644	351	433	227	46	24	10
hard/SFT	890	475	396	348	275	25	25	25
hard/RL_train	13,175	5,631	3,011	3,756	1,928	406	208	84
Total (**)	35,569	18,929	15,690	13,793	9,434	1,000	1,000	1,000

*If a puzzle has multiple sub-questions, the question types of these sub-questions will be separately counted.

**SFT training samples are selected from other training splits and thus are not included in the overall total.

🔧 Data Schema

Each JSONL entry contains the following fields:

{
  "problem": "string",      // Problem statement in Chinese
  "answer": "string",       // Answer. If multiple queries exist, answers will be separated by '===='
  "parameters": {           // Problem parameters
    "cond_num": "int",      // Number of conditions
    "sym_num": "int",       // Number of symbols
    "sym_type": "array",    // Types of symbols used
    "vars_scale": "float"   // Variable scaling factor (0-1 scale)
  },
  "config": {...},          // Problem configuration
  "qtype": "string",        // Question types ("简答题" | "单选题" | "填空题"). If multiple queries exist, their types will be separated by ','
  "eval_type": "string",    // Evaluation types ("numeral" | "nominal" | "ordered_array" | "unordered_array" | "ooa" | "oua"). If multiple queries exist, their types will be separated by ','
  "source": "string",       // Source seed puzzle identifier (e.g., "A1-ant")
  "id": "string",           // Unique problem identifier
  "difficulty": "float"     // Difficulty score (0-1 scale)
}

🧠 Baseline Model Performance

The following table summarizes the performance of several baseline models evaluated on the PuzzleCloneData test set. Metrics such as accuracy and F1 score are reported for each model.

Model	Normal	Hard	Avg.
Proprietary Models
ChatGPT-4o	31.6	23.9	27.7
ChatGPT-o3	86.8	82.0	84.4
Gemini-2.0-flash	41.7	30.5	36.1
Gemini-2.5-pro	75.8	67.2	71.5
Claude-3.5-sonnet	37.4	26.5	32.0
Claude-4-sonnet	62.2	46.3	54.3
GLM Series
GLM-Z1-9B-0414	63.0	51.9	57.5
GLM-Z1-32B-0414	70.5	59.0	64.8
Qwen2.5 Series
Qwen2.5-7B-Instruct	16.8	11.9	14.4
Qwen2.5-14B-Instruct	24.3	17.4	20.9
Qwen2.5-32B-Instruct	31.1	22.9	27.0
Qwen2.5-72B-Instruct	32.6	24.6	28.6
Qwen3 Series
Qwen3-8B	71.0	57.7	64.4
Qwen3-14B	78.1	65.0	71.6
Qwen3-32B	76.7	66.3	71.5
Qwen3-235B-A22B	82.0	71.2	76.6
DeepSeek Series
DeepSeek-R1-Distill-Qwen-14B	47.5	37.2	42.4
DeepSeek-R1-Distill-Qwen-32B	52.9	41.9	47.4
DeepSeek-R1-0528-Qwen3-8B	75.4	64.7	70.1
DeepSeek-R1-0528	87.7	80.2	84.0

📄 License

Please refer to the repository license for usage terms and conditions.

🤝 Citation

If you use this dataset in your research, please cite:

@misc{xiong2025puzzleclonesmtpoweredframeworksynthesizing,
      title={PuzzleClone: An SMT-Powered Framework for Synthesizing Verifiable Data}, 
      author={Kai Xiong and Yanwei Huang and Rongjunchen Zhang and Kun Chen and Haipang Wu},
      year={2025},
      eprint={2508.15180},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2508.15180}, 
}

📞 Contact

For questions, issues, or contributions, please open an issue on the GitHub repository.