datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
AI-CUDA-Engineer-Archive
The AI CUDA Engineer Archive 👷: Agentic CUDA Kernel Discovery, Optimization & Composition
We release The AI CUDA Engineer archive, a dataset consisting of approximately 30,000 CUDA kernels generated by The AI CUDA Engineer. It is released under the CC-By-4.0 license and can be accessed via HuggingFace and interactively visualized here. The dataset is based on the Kernel tasks provided in KernelBench and includes a torch reference implementation, torch, NCU and Clang-tidy… See the full description on the dataset page: https://huggingface.co/datasets/SakanaAI/AI-CUDA-Engineer-Archive.cudacuda-runtime-bundle
MBFS CUDA Runtime Bundle
Self-contained, ABI-matched CUDA execution stacks for running
ONNX Runtime on Windows with the CUDA / TensorRT
execution providers. One folder per CUDA major version:
Stack
Folder
CUDA
cuDNN
ONNX Runtime EP
When to use
cu12
cuda_v12/
12.x
9
CUDA + TensorRT 10
sm_75+ GPU on driver R527+ (CUDA 12)
cu11
cuda_v11/
11.x
8
CUDA + TensorRT 8.6
Pascal/Volta (sm_60–sm_70), or any GPU on a driver capped at CUDA 11.x
Pick the stack matching your… See the full description on the dataset page: https://huggingface.co/datasets/MBFSAITeam/cuda-runtime-bundle.kernelbench-cuda-tracesdim58-cudaData-31cases
Dim58 CPU Data — 31 Cases
Dataset uploaded from:
/mnt/data/ubuntu/research/outputs/data_cuda_geodesic58
Dataset summary
Property
Value
Repository
hosseinbv/dim58-cudaData-31cases
Number of files
64
Total size
17.99 GB
Source folder
data_cuda_geodesic58
File types
Extension
File count
.npz
62
.json
1
.csv
1
Top-level contents
0000_internal_case1_data.npz
0001_internal_B_10.npz… See the full description on the dataset page: https://huggingface.co/datasets/hosseinbv/dim58-cudaData-31cases.CUDA-Agent-Ops-6K
CUDA-Agent-Ops-6K
CUDA-Agent-Ops-6K is a curated training dataset for CUDA kernel generation and optimization.
It is released as part of the CUDA-Agent project:
Project Page: https://CUDA-Agent.github.io/
Github Repo: https://github.com/BytedTsinghua-SIA/CUDA-Agent
Dataset Summary
CUDA-Agent-Ops-6K contains 6,000 synthesized operator-level training tasks designed for large-scale agentic RL training. It is intended to provide diverse and executable CUDA-oriented training… See the full description on the dataset page: https://huggingface.co/datasets/BytedTsinghua-SIA/CUDA-Agent-Ops-6K.CUDA-L2
CUDA-L2: Surpassing cuBLAS Performance for Matrix Multiplication through Reinforcement Learning
🥳 Introduction
CUDA-L2 is a system that combines large language models (LLMs) and reinforcement learning (RL) to automatically optimize Half-precision General Matrix Multiply (HGEMM) CUDA kernels. CUDA-L2 systematically outperforms major matmul baselines to date, from the widely-used torch.matmul to state-of-the-art NVIDIA closed-source libraries (cuBLAS… See the full description on the dataset page: https://huggingface.co/datasets/ornith-ai/CUDA-L2.thick-py38-cuda11-env
THICK Python 3.8 CUDA 11 environment
This is a conda-pack archive of the environment used to reproduce THICK
MiniHack and VisualPinPad experiments.
Linux x86-64
Python 3.8.20
TensorFlow GPU 2.4.2
CUDA Toolkit 11.0.221
cuDNN 8.9.2.26
NumPy 1.19.5
Gym 0.17.2
MiniHack 0.1.5
NLE 0.9.0
W&B 0.15.10
Concatenate thick-py38-cuda11-linux-x86_64.tar.gz.part-* in lexical order,
verify the resulting archive with SHA256SUMS, extract it into a new Conda
environment directory, and run… See the full description on the dataset page: https://huggingface.co/datasets/datTrantien17/thick-py38-cuda11-env.MMD-Bench
MMD-Bench: Multimodal Model Degradation Benchmark
MMD-Bench is a comprehensive benchmark for evaluating vision-language models (VLMs) under various image degradation conditions. It covers 16 corruption types across 4 categories at 3 severity levels, applied to 6 widely-used VLM benchmarks.
Part of the CLEAR project.
[Paper] | [Code] | [Model]
Overview
Existing VLM benchmarks assume clean, high-quality images. In real-world scenarios, images often suffer from noise… See the full description on the dataset page: https://huggingface.co/datasets/CUDAOUTOFMEMORY/MMD-Bench.Nemotron-SFT-CUDA-v1
Dataset Description:
Nemotron-SFT-CUDA-v1 is a training dataset for CUDA code. It helps language models write CUDA kernels and solve CUDA programming problems.
We start from CUDA code in Nemotron Pretraining Code v2, which has a permissive license. An OpenCode agent powered by GLM-4.7 reads that code and writes new CUDA programming problems. Each problem comes with a hidden answer and tests. A second OpenCode + GLM-4.7 agent then tries to solve the problems using only the prompt… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Nemotron-SFT-CUDA-v1.video-editor-cudaCUDA-L1
| 🏠 Project Page | 📄 Paper | 🔥 Demo
🥳 Introduction
In this paper, we introduce CUDA-L1, an automated reinforcement learning (RL) framework for CUDA optimization. The core of CUDA-L1 is a contrastive RL model, a newly-designed RL system to enhance optimization through comparative learning.… See the full description on the dataset page: https://huggingface.co/datasets/ornith-ai/CUDA-L1.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.cudaLLM-data
CudaLLM Dataset
A high-quality dataset of PyTorch operator test cases, designed to benchmark and evaluate the capabilities of LLMs in generating optimized CUDA kernels. This dataset provides pairs of problems (standard PyTorch nn.Module implementations) and solutions (performance-optimized versions using custom CUDA kernels). It's a valuable resource for research in AI for HPC, code generation, and compiler optimization. The data is generated by DeepSeek R1, DeepSeel Coder-7B, and… See the full description on the dataset page: https://huggingface.co/datasets/ByteDance-Seed/cudaLLM-data.r8-eval-suite-5bucket
⚠️ CRITICAL: Ollama Inference Flag Required for derived models
If you train or serve any Qwen3.5-9B-derived model from this lineage via Ollama,
you MUST pass "think": false in /api/chat requests for chat / instruction following / tool use.
The qwen3.5 RENDERER auto-injects <think> tags causing 25-46% empty-answer rates without this flag.
See dataset cudabenchmarktest/r9-research-framework/_OLLAMA_INFERENCE_WARNING.md for the full lesson learned.
R8/R9 Five-Bucket… See the full description on the dataset page: https://huggingface.co/datasets/cudabenchmarktest/r8-eval-suite-5bucket.asm_cuda_to_amdmaple-preview-cuda-benchmarks
Maple Preview TQ2_0 CUDA Benchmarks
Reproducibility data for the TQ2_0 CUDA patches in
PascalAI2024/maple-preview-windows-cuda.
This repository contains benchmark data, patch files, hashes, and raw validation
evidence. It does not duplicate the Maple model weights.
Result
The fresh local A/B/B/A validation on an RTX 4080 SUPER reproduced the fused-MMQ
prompt-processing gain:
Variant
pp512 mean
pp512 median
tg128 mean
tg128 median
Correctness
MMQ enabled… See the full description on the dataset page: https://huggingface.co/datasets/x0me/maple-preview-cuda-benchmarks.r9-research-framework
R9 Research Framework — Qwen3.5-9B Distillation
⚠️ CRITICAL: READ FIRST — Ollama Inference Flag Required
If you serve any Qwen3.5-derived model from this lineage via Ollama,
you MUST pass "think": false in the /api/chat request body.
curl -X POST http://localhost:11434/api/chat \
-d '{"model": "qwen3.5-9b-r10:q4km", "think": false, "messages": [...], "stream": false}'
Without this flag the model will appear to "loop" and produce empty answers
on 25-46% of requests.… See the full description on the dataset page: https://huggingface.co/datasets/cudabenchmarktest/r9-research-framework.stable_diffusion_cpp_CUDATESTcuda-nsys-training
Qwythos Nsight Systems Profiling Agent Dataset
Multi-turn GPU profiling agent trajectories for fine-tuning Qwythos-9B (and similar tool-calling models) on NVIDIA Nsight Systems (nsys) + CUDA-L1 / KernelBench workloads.
Generated autonomously on an RTX 5090 by the model itself driving real profiling tools for ~33 hours.
Code: ai-hpc/prof-dataset-gen
Stats
Split
Rows
Notes
train
5,884
Accepted episodes (quality ≥ 0.55)
eval
309
5% holdout from accepted… See the full description on the dataset page: https://huggingface.co/datasets/gittensor-model-hub/cuda-nsys-training.cuda-error-resolution-analysisr8-thinking-fix-sft
⚠️ CRITICAL: Ollama Inference Flag Required for derived models
If you train or serve any Qwen3.5-9B-derived model from this lineage via Ollama,
you MUST pass "think": false in /api/chat requests for chat / instruction following / tool use.
The qwen3.5 RENDERER auto-injects <think> tags causing 25-46% empty-answer rates without this flag.
See dataset cudabenchmarktest/r9-research-framework/_OLLAMA_INFERENCE_WARNING.md for the full lesson learned.
R8 Thinking-Fix SFT… See the full description on the dataset page: https://huggingface.co/datasets/cudabenchmarktest/r8-thinking-fix-sft.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.router-bench
router-bench
Open corpus for Gittensor-TinyRouter
by James-Cuda.
Three milestones
Folder
Product goal
Predict / run
Needs
milestone1/
Prompt triage
domain + difficulty
CPU/GPU classifier; no API
milestone2/
Model↔prompt scoring / difficulty routing
which model (from scores or difficulty)
GPU optional; no live API
milestone3/
Full TinyRouter
3 models × 3 roles (Thinker/Worker/Verifier)
GPU + OPENROUTER_API_KEY
You control how many domain labels… See the full description on the dataset page: https://huggingface.co/datasets/James-Cuda/router-bench.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.dots-tts-runtime-cuda
dots.tts runtime payload
This private Kaggle dataset is generated by Phorcys.Tools.DotsTtsRuntimeUploader for PHRunner.Kaggle.Service.DotsTTS.
Runtime flavor: LinuxCuda
Python tag: python3.12
Generated UTC: 2026-09-10T00:49:22.3335893+00:00
The dataset intentionally contains runtime artifacts, not the GGUF model repository by default. Keep the model files in a separate private Kaggle dataset, for example kaggle-pool-account/dots-tts-v1-5-gguf.
Top-level runtime directories are… See the full description on the dataset page: https://huggingface.co/datasets/stokiz/dots-tts-runtime-cuda.marin-starcoderdata_cudacuda-challenge-flux-dump-holdout
FLUX.1-schnell activation & weight dump (holdout)
Holdout set for the W4A4 Quantized GEMM CUDA Challenge. Same layer / step / shape / prompt as the public dump, different seed — used to validate that submitted kernels generalize rather than hard-coding to specific activation values.
Seed: 42. Otherwise identical extraction to the public dump.
Files
weights.pt — bit-identical to the public dump (weights are seed-independent)
activations_1024x1024.pt — different from… See the full description on the dataset page: https://huggingface.co/datasets/nunchaku-ai/cuda-challenge-flux-dump-holdout.llama-cpp-cuda130-blackwell120cuda_amd