GPUMODE/KernelBook

Overview

dataset_permissive{.json/.parquet} is a curated collection of pairs of pytorch programs and equivalent triton code (generated by torch inductor) which can be used to train models to translate pytorch code to triton code. The triton code was generated using PyTorch 2.5.0 so for best results during evaluation / running the triton code we recommend using that version of pytorch.

Dataset Creation

The dataset was created through the following process:

1. Repository Collection: PyTorch repositories were collected from GitHub using repositories (and associated hashes) from the Stack v1.
2. PyTorch Module Extraction: We extracted the pytorch code from the repositories, and seperated them into individual torch.nn modules with appropriate dependencies.
3. Creating Unit Tests: We created unit tests for each module to ensure that the code was working as expected. Code in which could not create unit tests for was removed.
4. Extracting Triton Code: We used torch.compile in order to produce triton code from the pytorch code.
5. Transorming Triton Code: We transformed the triton code into one which resembled the format seen in KernelBench.
6. Metadata Enrichment: Each repository entry was enriched with metadata such as license information, star count, and commit SHA.

The scripts to do this yourself can be found here

Data Structure

Each entry in the dataset contains the following fields:

Field	Description
repo_name	The name of the repository in the format username/repository
licenses	List of licenses associated with the repository
stars	Number of GitHub stars the repository has
sha	The commit SHA hash used for version reference
repo_link	Direct link to the repository at the specific commit (GitHub URL)
Additional fields	The dataset may contain other repository-specific information

File Formats

The dataset is available in two formats:

1. JSON: dataset_permissive.json - A human-readable format that can be easily parsed by most programming languages.
2. Parquet: dataset_permissive.parquet - A columnar storage format optimized for analytics and big data processing.

Usage Examples

Loading the Dataset in Python

Using JSON:

import json

# Load the JSON version
with open('dataset_permissive.json', 'r') as f:
    repos = json.load(f)

# Example: Print the first 5 repository names
for repo in repos[:5]:
    print(repo['repo_name'])

Using Parquet:

import pandas as pd

# Load the Parquet version
df = pd.read_parquet('dataset_permissive.parquet')

# Example: Get repositories with more than 1000 stars
popular_repos = df[df['stars'] > 1000]
print(f"Number of popular repositories: {len(popular_repos)}")

License Information

The dataset_permissive contains only repositories with permissive licenses, including but not limited to:

• MIT License
• Apache License 2.0
• BSD Licenses (various)
• Mozilla Public License
• Unlicense
• zlib License

The dataset itself is provided for research and development purposes. Users should still verify the license of individual repositories before using their code in production or commercial settings.

Citation

@software{kernelbook2025,
    title={KernelBook},
    author={Paliskara, Sahan and Saroufim, Mark},
    year={2025},
    month={5},
    url={https://huggingface.co/datasets/GPUMODE/KernelBook},
}