Nemotron-SFT-CUDA-v1: Synthetic CUDA Programming Problems for Model Fine-Tuning

Nemotron-SFT-CUDA-v1 is a training dataset for CUDA code, created by NVIDIA. It contains synthetic CUDA programming problems and solutions generated by AI agents based on permissively licensed source code from the Nemotron Pretraining Code v2 dataset. The dataset was last updated on June 4, 2026.

Use Cases

• Fine-tuning language models to write CUDA kernels based on the described programming problems.
• Training models to solve CUDA programming problems using the hidden answers and tests.
• Benchmarking code generation agents on CUDA-specific tasks.
• Studying synthetic data generation methods for programming education.

Strengths

• Created by NVIDIA, a leading authority in GPU computing.
• Based on a permissively licensed source code corpus (Nemotron Pretraining Code v2).
• Includes hidden answers and tests for each generated problem.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data is synthetically generated by AI agents, which may introduce patterns not found in human-written code.

Provenance

Source

NVIDIA

Collection Method

Generated by OpenCode agents powered by GLM-4.7, which read CUDA code from Nemotron Pretraining Code v2 to write new problems and then attempt to solve them.

Freshness

Last updated 2026-06-04 04:35:43; freshness should be verified.