datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
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.gpu-prices
GPU Price Tracker
A continuously-updated dataset of cross-cloud GPU rental pricing
covering 13 public cloud providers (AWS, GCP, Azure, Lambda Labs,
RunPod, Vast.ai, DataCrunch, Cudo Compute, TensorDock, Vultr, Oracle,
Nebius, CloudRift): 3M+ listing observations, 70+ GPU types, collected
twice daily since January 2026 by scraping provider pricing surfaces via
the gpuhunt library and
published as Hive-partitioned Parquet files
(prices/dt=YYYY-MM-DD/*.parquet).
The dataset is… See the full description on the dataset page: https://huggingface.co/datasets/afhubbard/gpu-prices.backendbench_tests
TorchBench
The TorchBench suite of BackendBench is designed to mimic real-world use cases. It provides operators and inputs derived from 155 model traces found in TIMM (67), Hugging Face Transformers (45), and TorchBench (43). (These are also the models PyTorch developers use to validate performance.) You can view the origin of these traces by switching the subset in the dataset viewer to ops_traces_models and torchbench for the full dataset.
When running BackendBench, much of the… See the full description on the dataset page: https://huggingface.co/datasets/GPUMODE/backendbench_tests.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.l4-gpu-llm-benchmark-leaderboard
🚀 Local LLM Serving & Quality Benchmark Leaderboard (NVIDIA L4 24GB)
An exhaustive, reproducible benchmark study measuring real-world serving performance (TTFT, TPOT, throughput, peak VRAM, energy consumption, and cost) alongside rigorous task quality gates (HumanEval+, MMLU-Pro, BFCL v4 tool calling, and RULER needle retrieval) for open-weight LLMs on a single NVIDIA L4 24GB GPU.
📊 Executive Summary & Key Takeaways
⚡ Best Throughput & Coding Workhorse:… See the full description on the dataset page: https://huggingface.co/datasets/mayank-dubey-ai/l4-gpu-llm-benchmark-leaderboard.gpu-database
GPU Database
Comprehensive GPU specifications database with architecture, manufacturing, API support, performance details, and kernel development specs.
2,824 GPUs across NVIDIA, AMD, and Intel
Part of RightNow — AI-powered code editor for GPU kernel development
Data
Vendor
GPUs
File
NVIDIA
1,286
data/nvidia/all.json
AMD
1,292
data/amd/all.json
Intel
180
data/intel/all.json
All
2,824
data/all-gpus.json
Schema
Each GPU contains up to 55… See the full description on the dataset page: https://huggingface.co/datasets/Jr23xd23/gpu-database.triton-gpu-latency
Triton GPU Latency Dataset
A large dataset of PyTorch problems (mostly from KernelBench) paired with candidate Triton-kernel implementations and their measured GPU runtimes generated by MakoraGenerate. Each row is a self-contained Python program that defines (1) a reference Model written with plain PyTorch ops and (2) a ModelNew that re-implements the same forward pass with a hand-written or generated Triton kernel. The label is the runtime of executing ModelNew.
Built for… See the full description on the dataset page: https://huggingface.co/datasets/makora-ai/triton-gpu-latency.gpuark-gpu-dataset
GPU Ark — open GPU specifications & benchmarks dataset
Specifications of 13,566 GPUs released between 1999 and 2025 — from the GeForce 256 to
NVIDIA Blackwell and AMD Instinct MI355X — plus 993 third-party benchmark results.
Curated and maintained by GPU Ark (a GPU catalog & price comparison
project). Canonical source and always-fresh copy: https://gpuark.com/datasets/.
Files
File
Rows
What
gpuark-gpu-specs.csv
13,566
One row per GPU — public spec columns… See the full description on the dataset page: https://huggingface.co/datasets/Intelion/gpuark-gpu-dataset.one-layer-deeper-submissions
One Layer Deeper submissions
This dataset archives 15,602 accepted uploads from 206 GitHub accounts to the One Layer Deeper competition. It contains 9,627 distinct source files, all upload metadata, and stored evaluation results. Snapshot: September 7, 2026, 21:48 UTC, after the August 31 submission deadline.
Split
Uploads
Succeeded
Failed
easy
11,961
11,112
849
medium
2,704
2,509
195
hard
937
847
90
All accepted uploads are included: practice runs, failures… See the full description on the dataset page: https://huggingface.co/datasets/GPUMODE/one-layer-deeper-submissions.gpu-compatibility
Self-Hosted AI — GPU Compatibility, Recipes and Catalogue
Which open-weight AI models actually run on which consumer GPU, and what it takes
to get them running. 2 646 model×GPU verdicts across 98 models
and 27 cards, plus 973 full setup guides
(17 MB of markdown) written against specific hardware.
This is the machine-readable form of smeltcore.com. Every row carries a
url back to the page it came from.
Generated 2026-08-31T19:42:36+00:00 from the public read API… See the full description on the dataset page: https://huggingface.co/datasets/Smeltcore/gpu-compatibility.gpu-forecasters-puct-search-eventsCompanion artifact for GPU Forecasters: Language Models as Selective Surrogates for Kernel Runtime Optimization. Code: codezakh/gpu-surrogates.
Events emitted during the kernel searches in this work. You can reconstruct everything that happened in a search from these events: which kernels were tried, in what order, with what runtimes.
See the kernel-search code at codezakh/gpu-surrogates for the payload schema of each event kind.
Loading
from datasets import load_dataset
#… See the full description on the dataset page: https://huggingface.co/datasets/codezakh/gpu-forecasters-puct-search-events.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.hcc-array-supplementcloud-gpu-price-index
Cloud GPU Price Index
Canonical source: https://gpueconomy.com/price-index. That page is
recomputed every hour; this record is a dated snapshot of it, version
2026-09-19, built from data updated 2026-09-19T09:26:26.511913+00:00. When you cite,
cite GPU Economy and link the page; the snapshot is here so that a number you
used keeps existing exactly as you used it.
The index is the weekly median publicly listed on-demand price of one NVIDIA
H100 SXM GPU-hour across the cloud GPU… See the full description on the dataset page: https://huggingface.co/datasets/gpueconomy/cloud-gpu-price-index.gpu-forecasters-eval-set-predictionsCompanion artifact for GPU Forecasters: Language Models as Selective Surrogates for Kernel Runtime Optimization. Code: codezakh/gpu-surrogates.
Surrogate predictions on the held-out evaluation set. Each row is one forecast from one (surrogate, repeat) on one row of codezakh/gpu-forecasters-eval-set.
Loading
from datasets import load_dataset
# all surrogates and repeats
ds = load_dataset("codezakh/gpu-forecasters-eval-set-predictions", name="combined", split="predictions")
# one… See the full description on the dataset page: https://huggingface.co/datasets/codezakh/gpu-forecasters-eval-set-predictions.gpu-forecasters-eval-setCompanion artifact for GPU Forecasters: Language Models as Selective Surrogates for Kernel Runtime Optimization. Code: codezakh/gpu-surrogates.
Held-out evaluation set used in the paper. Each row is one (reference, candidate) kernel pair on a GPU Mode task, with the candidate's measured speedup over the reference on an A100.
Loading
from datasets import load_dataset
# all six packs combined
ds = load_dataset("codezakh/gpu-forecasters-eval-set", name="combined", split="eval")
#… See the full description on the dataset page: https://huggingface.co/datasets/codezakh/gpu-forecasters-eval-set.categorized_triton_data_permissive
Dataset Description
This dataset contains code snippets from Triton-based projects across GitHub, specifically filtered to include only repositories with permissive licenses (MIT, Apache, BSD, etc.). Each entry in the dataset includes:
Triton code snippet
Repository information
File path
Commit hash
Direct GitHub URL to the source code
License information
Categorization of the code functionality
Dataset Creation
The dataset was created by:
Collecting Triton code… See the full description on the dataset page: https://huggingface.co/datasets/GPUMODE/categorized_triton_data_permissive.gpu-mode-triton-augment-sample
Dataset Card for gpu-mode-triton-augment-sample
This dataset has been created with distilabel.
Dataset Summary
This dataset contains a pipeline.yaml which can be used to reproduce the pipeline that generated it in distilabel using the distilabel CLI:
distilabel pipeline run --config "https://huggingface.co/datasets/winglian/gpu-mode-triton-augment-sample/raw/main/pipeline.yaml"
or explore the configuration:
distilabel pipeline info --config… See the full description on the dataset page: https://huggingface.co/datasets/winglian/gpu-mode-triton-augment-sample.GPU-Resources-Estimation-for-Deep-Learning-Training-Tasks
GPUMemNet and GPUUtilNet Dataset
This dataset accompanies the paper
“GPU Memory and Utilization Estimation for Training-Aware Resource
Management: Opportunities and Limitations.”
It contains synthetic deep learning training configurations and their measured
GPU memory consumption and utilization characteristics.
Dataset configurations
The dataset is divided into separate configurations because MLP, CNN, and
Transformer workloads use different feature schemas.… See the full description on the dataset page: https://huggingface.co/datasets/ehyo/GPU-Resources-Estimation-for-Deep-Learning-Training-Tasks.autonomous-cloud-gpu-slurm-serving-suite
⚡ Autonomous Cloud GPU Infrastructure, Slurm Orchestration & Distributed Serving Suite (2026)
A Production-Grade, Verifiable Synthetic Corpus for Training Autonomous AI Supercomputing & LLM Serving Agents
⚡ Overview & Industry Problem
Operating massive AI supercomputers (thousands of NVIDIA H100/H200 and Blackwell GPUs) requires coordinating Slurm cluster schedules, topology-aware NVLink cliques, NCCL AllReduce rings, RoCE v2 lossless fabrics… See the full description on the dataset page: https://huggingface.co/datasets/beatsprom/autonomous-cloud-gpu-slurm-serving-suite.benchmark-dataset-different-gpu-workload
GPU catalog × LLM workload VRAM benchmark
Summary
Tabular benchmark in CSV form: each row pairs a catalog GPU (gpu_id, gpu_display_name, catalog_gpu_vram_gb) with a concrete LLM inference-style workload (model, parameter count, context length, precision, batch size, concurrent users). The file records math_engine VRAM component estimates (weights, KV cache, activations, overhead, totals, tier), a document_engine recommended VRAM value, a short comparison summary… See the full description on the dataset page: https://huggingface.co/datasets/odyn-network/benchmark-dataset-different-gpu-workload.ia-loaded-embedded-gpu
Dataset Card for "ia-loaded-embedded-gpu"
More Information needed
gpu-forecasters-rl-training-poolCompanion artifact for GPU Forecasters: Language Models as Selective Surrogates for Kernel Runtime Optimization. Code: codezakh/gpu-surrogates.
Training data used to fine-tune the three RL surrogate variants in the paper (correctness, correctness + Brier, correctness + CRPS rewards). Each row is one (anchor, candidate) kernel pair with the candidate's measured speedup relative to the anchor.
No rows for the gdn_chunk_fwd_o pack.
Loading
from datasets import load_dataset
ds =… See the full description on the dataset page: https://huggingface.co/datasets/codezakh/gpu-forecasters-rl-training-pool.yi-ko-DPO-noprompt-datasetGPU-Demand
Runpod GPU Weekly Snapshots: Current: 2026-05-24 → 2026-06-13
This package contains raw and cleaned exports of Runpod GPU snapshot data collected each week.
This dataset will be updated once a week on Sundays.
Newest time range
2026-05-24 → 2026-06-13
Coverage
Raw export: 41,930 rows across 449 timestamps
Cleaned export: 35,626 rows across 379 complete timestamps
Expected complete snapshot size: 94 rows per timestamp
Removed incomplete timestamps:… See the full description on the dataset page: https://huggingface.co/datasets/dejavoor/GPU-Demand.gpuemu-corpus
gpuemu Kernel-Correctness Corpus + Reproducibility Artifact
The 26-op corpus, experiment drivers, and analysis scripts behind the four
gpuemu preprints, led by "The Correctness Illusion in LLM-Generated GPU
Kernels" (arXiv:2606.20128). A
controlled set for measuring whether a correctness oracle actually catches the
bugs LLM-generated GPU kernels routinely contain — plus the full harness that
produces every table and figure in the papers.
Papers it backs
P1 — The… See the full description on the dataset page: https://huggingface.co/datasets/dipankarsarkar/gpuemu-corpus.energy_consumption_by_model_and_gpupallasbench-robust-gpu-a100
PallasBench: Robust Pallas GPU Kernel Benchmark (A100)
39/45 kernels passing on NVIDIA A100 80GB -- the first GPU-focused evaluation of JAX Pallas kernels.
What is this?
PallasBench is a suite of 45 JAX Pallas kernels across 3 difficulty levels. The original kernels were designed for TPU and failed on GPU because Pallas compiles to Triton on NVIDIA hardware, which has strict block size limits that TPU's Mosaic compiler does not.
We fixed all 45 kernels for GPU… See the full description on the dataset page: https://huggingface.co/datasets/EvanOLeary/pallasbench-robust-gpu-a100.gpu-spot-rental-prices
GPU spot-rental prices
Weekly medians of vast.ai marketplace asks, normalised to USD per GPU-hour (dph_total / num_gpus),
for H100, H200, B200, B300, A100 — on-demand and interruptible tiers where listed.
Method: cheapest-500 asks per SKU, so the series is deliberately low-biased; min / p25 / median per tier.
Machine-generated from vast.ai's public marketplace API. No internal figures, no vendor quotes.
data/prices.csv — the series, one row per GPU x tier per collection date… See the full description on the dataset page: https://huggingface.co/datasets/thsysmfh/gpu-spot-rental-prices.orca-alpaca-gpt4-math
Dataset Card for "orca-alpaca-gpt4-math"
More Information needed