CudaLLM Data: PyTorch Operator Test Cases for CUDA Kernel Generation

ByteDance-Seed provides a dataset of PyTorch operator test cases designed to benchmark LLMs in generating optimized CUDA kernels. It contains pairs of standard PyTorch nn.Module implementations and their performance-optimized versions using custom CUDA kernels. The dataset was last updated on August 3, 2025.

Use Cases

• Benchmarking LLM code generation capabilities based on PyTorch operator test cases.
• Training models for CUDA kernel optimization based on problem-solution pairs.
• Researching compiler optimization techniques based on the provided code transformations.

Strengths

• Dataset is described as 'high-quality' in the provided description.
• Contains pairs of standard and performance-optimized code implementations.
• Generated by models including DeepSeek R1 and DeepSeek Coder-7B.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

ByteDance-Seed

Collection Method

Generated by AI models including DeepSeek R1 and DeepSeek Coder-7B.

Freshness

Last updated 2025-08-03 02:29:03; freshness should be verified.