Skip to content

Dataset · nvidia

NVIDIA SFT dataset for CUDA code generation

nvidia/Nemotron-SFT-CUDA-v1

Synthetic dataset of 1k-10k tasks with tests from NVIDIA for training language models to write CUDA kernels.

What's inside

The dataset contains between 1k and 10k English text examples for the text-generation task. The CUDA programming tasks were generated by an OpenCode agent based on GLM-4.7 over source code from the open Nemotron Pretraining Code v2 corpus, with each task provided with a hidden answer and tests. A second agent solved these tasks based solely on the prompt text without source code access.

What you can do with it

  • SFT fine-tuning of code LLMs for writing and optimizing CUDA kernels
  • Evaluating model capability to solve algorithmic CUDA tasks with test verification
Tasks
Text generation
Languages
English
Size
1K–10K rows
Downloads this month
846
Likes
23

How to load it

from datasets import load_dataset

ds = load_dataset("nvidia/Nemotron-SFT-CUDA-v1")