datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
kernelbench-samples
KernelBench Samples
Samples from experiments for KernelBench, described in our arxiv
Learn more about KernelBench from our
Paper
Github Repo
The samples are organized as such
baseline_eval (Section 4 Baseline)
repeated_sampling (Section 5.1.1 Repeated Sampling)
iterative_refinement (Section 5.1.2 Iterative Refinement of Generations)
Within each folder, we organize the results by /level/model/problem_{id}/sample_{id}.
The inner most .json file contains the generated kernel and… See the full description on the dataset page: https://huggingface.co/datasets/ScalingIntelligence/kernelbench-samples.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.open-license-corpus
PubText
Welcome to the Open License Corpus (OLC), a 228B token corpus for training permissively-licensed language models.
Disclaimer: OLC should not be considered a universally safe-to-use dataset. We encourage users of OLC to consult a legal professional on the suitability of each data source for their application.
Dataset Summary
Domain
Sources
Specific License
# BPE Tokens (in billions; GPT-NeoX tokenizer)
Legal
Case Law, Pile of Law (PD subset)
Public… See the full description on the dataset page: https://huggingface.co/datasets/kernelmachine/open-license-corpus.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.p2-etf-kernel-conditional-moment-resultsfpga_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.kernelascent-tasks
KernelAscent — public dev split
KernelAscent is a benchmark for recursive self-improvement (RSI): a model optimizes
the GPU kernels used to train itself, and we measure whether kernel-optimization
capability compounds across rounds. This is the public dev split, released for
self-benchmarking and research; the leaderboard is scored on a private held-out split.
Project & code: https://github.com/ahmd-mohsin/KernelAscent
Leaderboard & docs:… See the full description on the dataset page: https://huggingface.co/datasets/muahmed7338/kernelascent-tasks.lin-alg-kernels-corekernelbench-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.kernelbench-v3-problems
KernelBench-v3 — Problem Definitions
The full set of problem definitions for KernelBench-v3 — the previous-generation sweep (2026-02) covering 10 frontier models across 3 NVIDIA GPUs (RTX 3090, H100, B200), with 43–58 problems per GPU.
Companion datasets:
Infatoshi/kernelbench-v3-runs — 2071 eval rows + winning agent solutions
Infatoshi/kernelbench-hard-problems — the newer KernelBench-Hard suite (single-Blackwell, 7 problems, 12 models)
Live site: https://kernelbench.com/v3
Source… See the full description on the dataset page: https://huggingface.co/datasets/Infatoshi/kernelbench-v3-problems.kernelbench-rag-contentKernelBench-M
KernelBench-M
The measurement artifact for Measuring the Checker: Mutation Analysis for
GPU-Kernel Benchmark Oracles: the mutation operators, the verified CUDA
substrates they mutate, the kill witnesses, and the pipeline that produced
every number in the paper.
Layout
rules/ 124 mutation rules, six families (mutator.py loads all of them)
substrates/ 208 gate-verified CUDA implementations, one per KernelBench
problem: the mutation… See the full description on the dataset page: https://huggingface.co/datasets/Elfsong/KernelBench-M.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.datakernelbench
DataKernelBench
Can LLMs optimize database queries on GPUs?
DataKernelBench evaluates LLMs on a novel task: optimizing analytical database queries as GPU kernels. It first represents each SQL query as a validated PyTorch program called a TorchPlan. It then evaluates LLMs by asking them to optimize either the tensor-intensive core (core) or the full query implementation (full) using CUDA or Triton, with execution-guided repair. The benchmark covers all 22 TPC-H queries.
On TPC-H… See the full description on the dataset page: https://huggingface.co/datasets/kerneldf/datakernelbench.kernelascent
KernelAscent — public dev split
KernelAscent is a benchmark for recursive self-improvement (RSI): a model optimizes
the GPU kernels used to train itself, and we measure whether kernel-optimization
capability compounds across rounds. This is the public dev split, released for
self-benchmarking and research; the leaderboard is scored on a private held-out split.
Project & code: https://github.com/ahmd-mohsin/KernelAscent
Leaderboard & docs:… See the full description on the dataset page: https://huggingface.co/datasets/muahmed7338/kernelascent.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.cuda-kernel-engineering
CUDA Kernel Engineering — Portfolio
A hands-on CUDA kernel engineering portfolio built on an NVIDIA L4 GPU (GCP).
Covers the complete path from first kernel to research-backed hypotheses that were
empirically falsified, with Nsight Compute profiling evidence at every step.
Each project teaches a specific optimization, measures its impact against cuBLAS,
and documents both positive and negative results.
Hardware: NVIDIA L4 (sm_89, 300 GB/s, 23 GB GDDR6)Stack: CUDA 12.4 (nvcc) /… See the full description on the dataset page: https://huggingface.co/datasets/rtferraz/cuda-kernel-engineering.linux-kernel-commits-aireason-instruct
Linux Kernel Code Patches Dataset
High-quality Linux kernel commit patches for training code generation and understanding models.
Dataset Description
This dataset contains 144,089 curated Linux kernel commits with:
Commit messages (instruction)
Smart-extracted code context (input)
Unified diff patches (output)
Optional AI quality scores and reasoning
Dataset Variants
Variant
Examples
Description
super_ultra
206
AI-recommended commits (Reasoning… See the full description on the dataset page: https://huggingface.co/datasets/ewedubs/linux-kernel-commits-aireason-instruct.titans-memory-kernels
Titans Memory Kernel - Nova Lineage
Evolved memory kernels for the Titans Memory system.
Structure
nova_prime.safetensors - The Prime kernel (16M params, 4096 dim)
nova_v1.safetensors - Production-hardened V1 (identity preserved)
REGISTRY.jsonl - Lineage registry with metadata
variants/ - 137 generation snapshots from evolution
Parameters
Dimensions: 4096
Memory Size: ~65MB per kernel
Evolution: Hardened with decay=0.99999, lr=0.0001
Usage
from… See the full description on the dataset page: https://huggingface.co/datasets/ADAPT-Chase/titans-memory-kernels.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.kernelbench-hard-problems
KernelBench-Hard — Problem Definitions
The 7 problem definitions for KernelBench-Hard, a benchmark for autonomous LLM coding agents writing GPU kernels on a single Blackwell GPU (RTX PRO 6000, sm_120, CUDA 13.2).
Companion datasets:
Infatoshi/kernelbench-hard-runs — 84 agent transcripts, winning solutions, leaderboard, reward-hack annotations
Live site: https://kernelbench.com/hard
Methodology blog: https://kernelbench.com/blog/hard
Source repo:… See the full description on the dataset page: https://huggingface.co/datasets/Infatoshi/kernelbench-hard-problems.SemanticAlign-Bench
SemanticAlign-Bench
A benchmark for evaluating AI agents on structured claim extraction from top-tier ML conference papers. Each paper is decomposed into Semantic Alignment Units (SAU) — atomic, self-contained implementation propositions — across four diagnostic dimensions spanning numerical precision to pipeline-level workflow. Agents are evaluated on whether they can reproduce these claims without hallucination, omission, or misordering.
The Four SAU Dimensions… See the full description on the dataset page: https://huggingface.co/datasets/kernel-14/SemanticAlign-Bench.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.dr-kernel-RL