Skip to content

Dataset · nvidia

Synthetic reasoning corpus for code generation

nvidia/OpenCodeReasoning-2

2.5M Python and C++ code examples with reasoning traces from NVIDIA for SFT training.

What's inside

The largest synthetic reasoning dataset for programming at release time. Contains 1.4M Python examples and 1.1M C++ examples derived from 34,799 unique competitive programming problems. Data targets code completion and code critique tasks.

What you can do with it

  • Supervised fine-tuning (SFT) of language models for code generation and completion
  • Training critic models for bug detection and program analysis
  • Enhancing step-by-step logical reasoning skills in code LLMs
Tasks
Text generation
Size
1M–10M rows
Downloads this month
15.8K
Likes
64

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("nvidia/OpenCodeReasoning-2")