sungyub/deepmath-103k-verl

DeepMath-103K VERL

📊 Dataset Summary

This dataset contains 95,496 mathematical reasoning problems in VERL format, processed from zwhe99/DeepMath-103K.

Key Features:

• 95,496 high-quality math problems
• Converted to VERL format for reward modeling
• Verified ground truth answers
• Ready for reinforcement learning training

---

🔗 Source Dataset

Original Repository

• Repository: zwhe99/DeepMath-103K

• License: MIT

• Paper: DeepMath: Deep Learning for Mathematical Reasoning

• Authors: Zhenwen He et al.

Dataset Description

DeepMath-103K is a curated collection of 103,000 mathematical problems covering various difficulty levels and mathematical domains. The dataset emphasizes step-by-step reasoning and verification.

---

🔄 Preprocessing Pipeline

This dataset has been preprocessed and converted to the VERL (Verification and Reinforcement Learning) format for use in mathematical reasoning tasks with reward modeling.

Cleaning Methodology

Standard Processing:

• URL filtering (samples containing URLs removed)
• Format normalization to VERL schema
• Basic text cleaning and validation

Deduplication

Intra-dataset Deduplication:

• Method: SHA-256 hash-based with text normalization
• Before deduplication: 103,000 samples
• After deduplication: 95,496 samples
• Reduction: 7.3%

Inter-dataset Deduplication (v3.0):

• Priority level: 4
• Cross-dataset duplicates removed: 5,000

---

💡 Preprocessing Examples

Example 1: Standard Processing (No Cleaning Preset)

Before Cleaning:

Calculate the derivative of $f(x) = x^2 + 3x + 2$ with respect to $x$.

After Cleaning:

Calculate the derivative of $f(x) = x^2 + 3x + 2$ with respect to $x$.

Changes Applied:

• ✓ Format normalization to VERL schema
• ✓ URL filtering (samples with URLs removed)
• ✓ Basic text validation
• ✓ No artifact removal applied

---

📐 VERL Schema

This dataset follows the standardized VERL (Verification and Reinforcement Learning) format:

{
  "data_source": "openai/gsm8k",
  "prompt": [
    {
      "content": "Calculate the sum of all odd numbers from 1 to 99.",
      "role": "user"
    }
  ],
  "ability": "math",
  "reward_model": {
    "style": "rule",
    "ground_truth": "\boxed{2500}",
    "hash": "sha256:abc123..."
  },
  "extra_info": {
    "split": "train"
  }
}

Field Descriptions

Field	Type	Description
data_source	string	Original dataset identifier (e.g., openai/gsm8k, numina_aime)
prompt	list[dict]	User query in chat format with role and content
ability	string	Task type (always "math" for this dataset)
reward_model.style	string	Reward computation method ("rule" for rule-based verification)
reward_model.ground_truth	string	Expected answer for verification (often in \boxed{} format)
reward_model.hash	string	SHA-256 hash of prompt content for deduplication
extra_info.split	string	Original split identifier ("train", "test", etc.)

---

📈 Dataset Statistics

Sample Distribution

• Total Samples: 95,496
• Dataset Size: 10.5 MB
• Average Problem Length: N/A characters

Data Sources

Distribution of problems by original data source:

Source	Count	Percentage
Mixed Sources	95,496	100%

Note: Detailed source distribution statistics will be added in future updates.

---

🚀 Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("sungyub/deepmath-103k-verl")

# Load with streaming (recommended for large datasets)
dataset = load_dataset("sungyub/deepmath-103k-verl", streaming=True)

# Preview first few examples
for example in dataset['train'].take(5):
    print(example['prompt'][0]['content'])  # User question
    print(example['reward_model']['ground_truth'])  # Answer
    print("---")

Using with VERL

from datatrove.utils.reward_score import compute_score

# Compute reward score for a generated solution
score = compute_score(
    data_source=example['data_source'],
    solution_str=generated_solution,
    ground_truth=example['reward_model']['ground_truth'],
    format_type="auto"  # Auto-detect XML or GPT OSS format
)

print(f"Reward score: {score}")

Integration with DataTrove

from datatrove.pipeline.readers import ParquetReader
from datatrove.pipeline.filters import LambdaFilter
from datatrove.executor import LocalPipelineExecutor

pipeline = [
    ParquetReader("sungyub/deepmath-103k-verl", text_key="prompt"),
    LambdaFilter(lambda doc: len(doc.text) > 100),  # Filter short problems
    # Add more processing steps...
]

executor = LocalPipelineExecutor(pipeline=pipeline, tasks=4)
executor.run()

---

📚 Citation

Original Dataset

@article{he2025deepmath,
  title = {DeepMath: Deep Learning for Mathematical Reasoning},
  author = {He, Zhenwen and others},
  journal = {arXiv preprint arXiv:2504.11456},
  year = {2025},
  url = {https://arxiv.org/abs/2504.11456}
}

This Processed Version

@dataset{sungyub_math_verl_deepmath-103k-verl,
  author = {Sungyub Kim},
  title = DeepMath-103K VERL,
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/sungyub/deepmath-103k-verl}}
}

---

⚖️ License

• This processed dataset: mit
• Original dataset: MIT

---

🙏 Acknowledgments

This dataset was processed using the DataTrove library.

Credits:

• Original dataset authors: Zhenwen He et al.
• Processing and VERL conversion: Sungyub Kim
• MathDatasetCleaner implementation: DataTrove contributors

Special thanks to: Zhenwen He and collaborators for the DeepMath dataset

---

📝 Version History

v1.0.0 (Initial Release)

• Processed 95,496 samples from zwhe99/DeepMath-103K
• Converted to VERL format
• Standard processing applied
• Ready for reinforcement learning training

---

🔗 Related Resources

• Unified Collection: sungyub/math-verl-unified - All 9 math datasets with inter-dataset deduplication
• DataTrove Documentation: https://github.com/huggingface/datatrove
• VERL Format Specification: See VERL Schema section above

---

Questions or issues? Open an issue on the DataTrove GitHub repository