KernelBench-Mega: Agent Traces for GPU Megakernel Performance

23 agent traces of coding agents writing full GPU megakernels for Blackwell, H100, and B200 architectures, scored for speedup over a reference. The dataset, created by Infatoshi, was last updated on June 20, 2026. It includes a manifest mapping runs to model, harness, problem, GPU, and score, and is contamination-audited.

Use Cases

• Analyze agent reasoning patterns for GPU kernel generation based on the Claude-Code session format traces.
• Benchmark model performance on GPU kernel coding tasks based on the speedup scores and hardware targets mentioned.
• Study potential contamination in code generation models based on the described audit of 23 verified cells.
• Compare code generation strategies across different models and GPU architectures based on the manifest metadata.

Strengths

• Contains 23 distinct, verified agent traces, providing a concrete sample size.
• Includes performance scores (speedup over reference) for generated GPU kernels.
• Manifest file provides structured metadata linking each trace to model, harness, problem, and GPU.
• Traces are contamination-audited, suggesting a focus on data quality for evaluation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the trace data is unknown, which may limit suitability assessment.
• The dataset description references an external page for full details, indicating incomplete on-platform metadata.

Provenance

Source

huggingface

Collection Method

Likely collected from agent sessions running coding tasks on the KernelBench platform.

Freshness

Last updated 2026-06-20 01:44:17; freshness should be verified.