OpenCodeReasoning-2: Python and C++ Samples for Code Critique

OpenCodeReasoning-2 contains 1.4 million Python and 1.1 million C++ samples derived from 34,799 unique competitive programming questions. This synthetic dataset is designed for supervised fine-tuning tasks in code completion and critique. It was created by NVIDIA and released on the Hugging Face platform in May 2025.

Use Cases

• Fine-tune a model for Python code completion using the 1.4M Python samples.
• Train a model to critique C++ code based on the 1.1M C++ reasoning samples.
• Benchmark reasoning capabilities of code generation models on 34,799 unique programming questions.
• Develop a model for generating code explanations or corrections using the synthetic reasoning data.

Strengths

• 2.5 million total samples across two programming languages.
• Derived from 34,799 unique competitive programming questions, providing diverse problem coverage.
• Dataset is explicitly designed for supervised fine-tuning of code-related tasks.

Limitations

• The dataset is synthetic, which may introduce patterns or biases not present in real-world code.
• Specific column structure, data quality metrics, and sample size details are not provided in the available description.
• The dataset's temporal coverage and update frequency are unknown.

Provenance

Source

NVIDIA via Hugging Face.

Collection Method

Synthetic generation pipeline, details referenced in a linked GitHub repository.

Time Range

null

Freshness

null

Geography

null