KernelBot: 100K+ Optimized AMD MI300 GPU Kernel Submissions

GPUMODE released this dataset in early 2026, containing between 100,000 and 1,000,000 GPU kernel submissions from the KernelBot competition platform. The collection focuses on optimized code specifically targeting AMD MI300 hardware and includes subsets for successful and deduplicated entries.

Use Cases

• Training Large Language Models on optimized AMD MI300 GPU kernel code
• Analyzing common coding patterns in successful vs. unsuccessful kernel submissions using the correctness status
• Benchmarking code deduplication algorithms using the deduplicated_submissions.parquet file

Strengths

• Contains between 100,000 and 1,000,000 records
• Includes a pre-filtered successful_submissions.parquet file for verified working code
• Provides deduplicated_submissions.parquet to minimize redundant code patterns

Limitations

• Hardware-specific focus on AMD MI300 may limit generalizability to NVIDIA or other architectures
• Lack of detailed performance metrics such as execution time or TFLOPS in the basic file descriptions

Provenance

Source

GPUMODE KernelBot competition platform

Collection Method

Aggregated submissions from a competitive coding platform

Time Range

Up to February 2026

Freshness

Last updated February 24, 2026.

Geography

United States