kcxain/KernelZero-CUDA-SFT

KernelZero CUDA SFT

KernelZero CUDA SFT contains the 43,302 examples used for the two-stage supervised cold start of the KernelZero CUDA Coder. The dataset accompanies KernelZero: Co-Evolving Proposer and Coder for Continuously Improved GPU Kernel Generation.

Dataset composition

Stage	Examples	Description
Stage 1	24,073	Initial CUDA reasoning SFT
Stage 2	19,229	Continued CUDA SFT with later validated generations
Total	43,302	Unique training examples

Every released example passed response-format checking, CUDA compilation, and numerical validation against its PyTorch reference on the original evaluation infrastructure.

Fields

Field	Type	Description
input	string	Exact SFT input prompt, including the target PyTorch module and CUDA optimization instructions
output	string	Exact SFT target response, including reasoning and the generated implementation
torch_module	string	Standalone reference PyTorch module
cuda_code	string	Generated CUDA/Python implementation with the outer Markdown code fence removed when present
speedup	float64	Measured speedup over the PyTorch reference on an NVIDIA A100-80GB GPU
operators	list[string]	Torch operators represented by the example
source_stage	string	SFT stage from which the example originates
source_index	int64	Original source index within the CudaLLM-derived corpus

Load the dataset

from datasets import load_dataset

dataset = load_dataset("kcxain/KernelZero-CUDA-SFT", split="train")
print(dataset[0])

Notes

The speedup field preserves the measurement stored with each accepted training example. Runtime measurements are hardware- and software-dependent. Reproduction should use the KernelZero validation environment and the same A100-80GB target.

Citation

@article{kernelzero2026,
  title={KernelZero: Co-Evolving Proposer and Coder for Continuously Improved GPU Kernel Generation},
  author={Ke, Changxin and Zhang, Rui and Fang, Zixiang and Li, Zhenghong and Wen, Yuanbo and Shen, Jiashuo and Wang, Shuo and Guo, Jiaming and Li, Ling and Guo, Qi and Chen, Yunji},
  journal={arXiv preprint arXiv:2609.33074},
  year={2026}
}