datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
kernelbench-mega-traces
KernelBench-Mega agent traces
Coding agents writing full GPU megakernels across Blackwell / H100 / B200, scored as speedup over reference; contamination-audited (23 verified cells).
Each .jsonl file is one agent run in Claude-Code session format, viewable with the agent trace viewer. Filename = run id; manifest.csv maps each run to model / harness / problem / GPU / score.
23 agent traces · live leaderboard: https://kernelbench.com/mega
Secrets redacted. Full reasoning for… See the full description on the dataset page: https://huggingface.co/datasets/Infatoshi/kernelbench-mega-traces.kernelbench-hard-traces
KernelBench-Hard agent traces
Frontier coding agents writing optimized CUDA/Triton kernels (FP8 GEMM, paged
attention, MoE, W4A16, KDA, Top-k) on RTX PRO 6000 Blackwell, H100 PCIe, and
B200; roofline-graded.
Each .jsonl file is one agent run in Claude-Code session format, viewable with
the Hugging Face Agent Trace viewer (Data Studio → open a row). Filename =
run id.
Live leaderboard: https://kernelbench.com/hard
Secrets redacted. Full reasoning for open-provider routes… See the full description on the dataset page: https://huggingface.co/datasets/Infatoshi/kernelbench-hard-traces.kernelbench-v3-runs
KernelBench-v3 — Agent Runs
2071 agent evaluations from the v3 sweep (2026-02): 10 frontier models × {RTX 3090, H100, B200} × 43–58 problems per GPU. Each row is one (model, gpu, problem) triple with correctness, speedup, baseline timing, token usage, cost, and a pointer to the agent's winning solution.py.
Companion datasets:
Infatoshi/kernelbench-v3-problems — 60 problem definitions
Infatoshi/kernelbench-hard-runs — newer KernelBench-Hard sweep (12 models × 7 problems on Blackwell… See the full description on the dataset page: https://huggingface.co/datasets/Infatoshi/kernelbench-v3-runs.kernelbench-cuda-tracesKernelBench
KernelBench
A benchmark designed to evaluate the ability of LLMs to generate efficient GPU kernels for optimizing neural network performance
Version
[07-21-2025] This HF dataset version has been updated to v0.1
Citation
@misc{ouyang2024kernelbench,
title={KernelBench: Can LLMs Write GPU Kernels?},
author={Anne Ouyang and Simon Guo and Azalia Mirhoseini},
year={2024},
url={https://scalingintelligence.stanford.edu/blogs/kernelbench/},
}
kernelbot-data
KernelBot Competition Data
This dataset contains GPU kernel submissions from the KernelBot competition platform. Submissions are optimized GPU kernels written for specific hardware targets.
Data Files
AMD MI300 Submissions
File
Description
submissions.parquet
All AMD competition submissions
successful_submissions.parquet
AMD submissions that passed correctness tests
deduplicated_submissions.parquet
AMD submissions deduplicated by… See the full description on the dataset page: https://huggingface.co/datasets/GPUMODE/kernelbot-data.fpga_cost_model_kernel_data
FPGA HLS Kernel Cost-Model Data
Evolved Vitis HLS C++ kernels paired with their ground-truth Vitis HLS
csynth results. Each row is one generated program from an evolutionary FPGA
optimisation run, linked to its kernel source, evaluator report.json, and raw
synthesis report.
Each row carries a split label: train marks the original benchmarks used
to fit the analytical cost model's learned correction term, and holdout marks
benchmarks added afterwards that were not used for… See the full description on the dataset page: https://huggingface.co/datasets/adimnaku/fpga_cost_model_kernel_data.kernelbench-hard-runs
KernelBench-Hard — Agent Runs
84 full agent transcripts (12 frontier models × 7 problems) from the KernelBench-Hard sweep on a single Blackwell GPU (RTX PRO 6000, sm_120, CUDA 13.2). Each run contains the model's full reasoning trace, every tool call, the final solution.py, and the eval result.
Companion datasets:
Infatoshi/kernelbench-hard-problems — the 7 problem definitions
Live site: https://kernelbench.com/hard
100 themed transcript viewers (HTML): https://kernelbench.com/runs… See the full description on the dataset page: https://huggingface.co/datasets/Infatoshi/kernelbench-hard-runs.QwQ-32B_enable-liger-kernel_False_OpenThoughts3_3k_eval_5554
mlfoundations-dev/QwQ-32B_enable-liger-kernel_False_OpenThoughts3_3k_eval_5554
Precomputed model outputs for evaluation.
Evaluation Results
Summary
Metric
AIME24
AMC23
MATH500
MMLUPro
JEEBench
GPQADiamond
LiveCodeBench
CodeElo
CodeForces
AIME25
HLE
LiveCodeBenchv5
HMMT
Accuracy
75.7
98.8
90.4
58.1
73.7
68.2
41.9
46.8
47.2
67.7
13.9
64.3
52.0
AIME24
Average Accuracy: 75.67% ± 1.57%
Number of Runs: 10
Run
Accuracy
Questions… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations-dev/QwQ-32B_enable-liger-kernel_False_OpenThoughts3_3k_eval_5554.QwQ-32B_enable-liger-kernel_False_OpenThoughts3_1k_eval_5554fpga_cost_model_kernel_data_attention_p2
FPGA HLS Kernel Cost-Model Data
Evolved Vitis HLS C++ kernels paired with their ground-truth Vitis HLS
csynth results. Each row is one generated program from an evolutionary FPGA
optimisation run, linked to its kernel source, evaluator report.json, and raw
synthesis report.
Each row carries a split label: train marks the original benchmarks used
to fit the analytical cost model's learned correction term, and holdout marks
benchmarks added afterwards that were not used for… See the full description on the dataset page: https://huggingface.co/datasets/adimnaku/fpga_cost_model_kernel_data_attention_p2.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:… See the full description on the dataset page: https://huggingface.co/datasets/GPUMODE/KernelBook.gspc-kernel-results
Kaggle 3090 ladder — prompt-bank passes
Prompt-bank passes from the Kaggle 3090 ladder. Each row of
kernel_results.jsonl carries the axis, the model family and full model id, the prompt, the score,
the n behind that score, a sigil content hash, the platform it ran on and the timestamp. Most rows are
n=1 single-prompt passes — read them as a ladder sweep across many open models, not as board n.
The live board is the authority
GET https://councilof.ai/api/gspc —… See the full description on the dataset page: https://huggingface.co/datasets/csoai/gspc-kernel-results.kern-kernels
kern-kernels
Reproducible attention kernel recipes, ABI manifests, checksums and measured
results. First profile: GB300 / Qwen3.8-27B / BF16 / Q24-KV4-D256 / page 64.
Uses unmodified TRTLLM-GEN full attention, not MLA. The model's GDN layers are
unchanged. Model weights and the base export's other kernels are not included.
NVIDIA binaries are downloaded directly from pinned upstream URLs and verified
by SHA256; this repository does not mirror them. The small Apache-2.0 vLLM KV… See the full description on the dataset page: https://huggingface.co/datasets/susun-123/kern-kernels.QwQ-32B_enable-liger-kernel_False_OpenThoughts3_10k_eval_5554
mlfoundations-dev/QwQ-32B_enable-liger-kernel_False_OpenThoughts3_10k_eval_5554
Precomputed model outputs for evaluation.
Evaluation Results
Summary
Metric
AIME24
AMC23
MATH500
MMLUPro
JEEBench
GPQADiamond
LiveCodeBench
CodeElo
CodeForces
AIME25
HLE
LiveCodeBenchv5
HMMT
Accuracy
77.7
98.8
91.2
55.3
67.5
63.5
41.6
47.1
48.1
70.0
13.5
65.0
49.0
AIME24
Average Accuracy: 77.67% ± 1.25%
Number of Runs: 10
Run
Accuracy
Questions… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations-dev/QwQ-32B_enable-liger-kernel_False_OpenThoughts3_10k_eval_5554.QwQ-32B_enable-liger-kernel_False_OpenThoughts3_10k_eval_8179
mlfoundations-dev/QwQ-32B_enable-liger-kernel_False_OpenThoughts3_10k_eval_8179
Precomputed model outputs for evaluation.
Evaluation Results
Summary
Metric
AIME24
AMC23
MATH500
JEEBench
GPQADiamond
LiveCodeBench
CodeElo
CodeForces
AIME25
HLE
LiveCodeBenchv5
HMMT
Accuracy
78.0
98.8
91.4
67.6
66.2
82.1
46.7
48.4
69.0
12.4
65.6
52.0
AIME24
Average Accuracy: 78.00% ± 1.71%
Number of Runs: 10
Run
Accuracy
Questions Solved
Total… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations-dev/QwQ-32B_enable-liger-kernel_False_OpenThoughts3_10k_eval_8179.cuda-triton-gpu-kernels-2026
⚡ Complete 2026 CUDA & OpenAI Triton High-Performance GPU Kernel Engineering SFT/DPO Suite
The definitive, production-grade synthetic alignment dataset engineered for training and fine-tuning open-weights Large Language Models (Qwen 2.5 Coder, DeepSeek-Coder, Llama 3.1) on ultra-high-throughput GPU kernel programming: NVIDIA Hopper H100 / Blackwell B200 TMA async transfers, OpenAI Triton 3.1+ FlashAttention-3, 32-bank conflict elimination, and low-bit FP8 / INT4 GEMM… See the full description on the dataset page: https://huggingface.co/datasets/beatsprom/cuda-triton-gpu-kernels-2026.glm-5.2-kernelgym-rollouts
GLM-5.2 KernelGym Rollouts
This dataset contains 3,200 feedback-driven GPU-kernel optimization trajectories
generated by zai-org/GLM-5.2-FP8: 100 validation tasks, two backends (inline
CUDA and Triton), and 16 rollouts per task.
Each trajectory retains the prompt/feedback message history, model responses and
reasoning, extracted kernel code, KernelGym compilation and correctness results,
profiling metadata, token usage, and stopping decision. Every published record
ended with… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/glm-5.2-kernelgym-rollouts.QwQ-32B_enable-liger-kernel_False_OpenThoughts3_3k_eval_8179
mlfoundations-dev/QwQ-32B_enable-liger-kernel_False_OpenThoughts3_3k_eval_8179
Precomputed model outputs for evaluation.
Evaluation Results
Summary
Metric
AIME24
AMC23
MATH500
JEEBench
GPQADiamond
LiveCodeBench
CodeElo
CodeForces
AIME25
HLE
LiveCodeBenchv5
HMMT
Accuracy
79.3
99.2
90.8
73.1
65.7
83.8
47.5
48.4
66.0
13.0
65.9
51.7
AIME24
Average Accuracy: 79.33% ± 1.23%
Number of Runs: 10
Run
Accuracy
Questions Solved
Total… See the full description on the dataset page: https://huggingface.co/datasets/mlfoundations-dev/QwQ-32B_enable-liger-kernel_False_OpenThoughts3_3k_eval_8179.autonomous-linux-kernel-ebpf-xdp-suite
⚡ Autonomous Linux Kernel, eBPF & XDP Programmable Dataplane Suite (2026)
A Production-Grade, Verifiable Synthetic Corpus for Training Autonomous Linux Kernel & eBPF Systems Agents
⚡ Overview & Industry Problem
Modern hyperscale cloud datacenters, bare-metal Kubernetes clusters, and low-latency financial trading nodes rely on in-kernel programmable dataplanes: eBPF, AF_XDP zero-copy rings, Traffic Control (TC) shapers, BPF LSM security hooks… See the full description on the dataset page: https://huggingface.co/datasets/beatsprom/autonomous-linux-kernel-ebpf-xdp-suite.kernel-vuln-dataset-full
Linux Kernel Vulnerability-Introducing Commits Dataset
Dataset Description
A labeled dataset of 1,426,202 Linux kernel git commits with full metadata, diffs, and binary labels indicating whether each commit introduced a vulnerability that was later fixed.
Intended use: Training and evaluating models for vulnerability-introducing commit detection — predicting whether a given code change will later require a security or bug fix.
How the Data Was Collected… See the full description on the dataset page: https://huggingface.co/datasets/quguanni/kernel-vuln-dataset-full.fpga_cost_model_kernel_data_mamba_p2
FPGA HLS Kernel Cost-Model Data
Evolved Vitis HLS C++ kernels paired with their ground-truth Vitis HLS
csynth results. Each row is one generated program from an evolutionary FPGA
optimisation run, linked to its kernel source, evaluator report.json, and raw
synthesis report.
Each row carries a split label: train marks the original benchmarks used
to fit the analytical cost model's learned correction term, and holdout marks
benchmarks added afterwards that were not used for… See the full description on the dataset page: https://huggingface.co/datasets/adimnaku/fpga_cost_model_kernel_data_mamba_p2.kernelbench_with_promptsThis is a version of KernelBench where the prompts to produce the Triton and cuda kernel are explicitly saved in the JSON data files.
It only contains Level 1, 2, 3 kernels.
The prompt is the same as what is provided in the original KernelBench repo.
The dataset is prepared by Jiin Woo during her internship at AWS Annapurna Labs, the lab behind Trainium chips.
This dataset is part of an unreleased paper, and the paper will be updated in this README soon. If you use this dataset, please cite… See the full description on the dataset page: https://huggingface.co/datasets/allenanie/kernelbench_with_prompts.kernel-vuln-dataset-full
Linux Kernel Vulnerability-Introducing Commits Dataset
Dataset Description
A labeled dataset of 1,426,202 Linux kernel git commits with full metadata, diffs, and binary labels indicating whether each commit introduced a vulnerability that was later fixed.
Intended use: Training and evaluating models for vulnerability-introducing commit detection — predicting whether a given code change will later require a security or bug fix.
How the Data Was Collected… See the full description on the dataset page: https://huggingface.co/datasets/pebblebed/kernel-vuln-dataset-full.ga104-cuda-kernels
GA104 Hand-Optimized CUDA Kernel Corpus
A measurement corpus of hand-optimized CUDA / SASS kernels targeting the
RTX 3070 Ti (GA104, sm_86, Ampere). Every kernel is written without
cuBLAS, cuDNN, or PyTorch in the optimized path; vendor libraries are
linked only for measured comparison under kernels/reference/. This
dataset is for SASS and GPU-optimization researchers — it pairs each
.cu source with its compiled machine code and its disassembly, so the
exact instruction stream a… See the full description on the dataset page: https://huggingface.co/datasets/pjt222/ga104-cuda-kernels.fpga_cost_model_kernel_data_qwen3_mlp
FPGA HLS Kernel Cost-Model Data
Evolved Vitis HLS C++ kernels paired with their ground-truth Vitis HLS
csynth results. Each row is one generated program from an evolutionary FPGA
optimisation run, linked to its kernel source, evaluator report.json, and raw
synthesis report.
Each row carries a split label: train marks the original benchmarks used
to fit the analytical cost model's learned correction term, and holdout marks
benchmarks added afterwards that were not used for… See the full description on the dataset page: https://huggingface.co/datasets/adimnaku/fpga_cost_model_kernel_data_qwen3_mlp.autonomous-gpu-kernel-triton-cuda-suite-2026
⚡ Autonomous GPU Kernel, Triton & CUDA Architecture Suite (2026)
A Production-Grade, Verifiable Synthetic Corpus for Training Frontier Coding Models (Qwen 3.8, DeepSeek-V3, Llama 3.3)
⚡ Overview & Industry Problem
Modern deep learning accelerators, custom ASICs, and high-performance computing clusters demand specialized, autonomous GPU kernel infrastructure: OpenAI Triton fused kernels, FlashAttention-3 forward/backward online softmax… See the full description on the dataset page: https://huggingface.co/datasets/beatsprom/autonomous-gpu-kernel-triton-cuda-suite-2026.kernelbook-glm4-evalsdaVinci-kernel-sftParallelKernelBench_Kernels
ParallelKernelBench Kernels
Net-new multi-GPU CUDA kernels generated by LLMs for ParallelKernelBench.
Each subdirectory under solutions/ is one model run. File names match the benchmark problem stems (e.g. 17_rope_allgather_cuda.py ↔ problem 17_rope_allgather in willychan21/ParallelKernelBench_Problems).
Layout
solutions/
<run_id>/
<stem>_cuda.py
...
Runs (1 run(s), 87 kernel files)
run_id
kernels
path… See the full description on the dataset page: https://huggingface.co/datasets/willychan21/ParallelKernelBench_Kernels.