Nemotron-Competitive-Programming-v2 is a large-scale synthetic programming and reasoning dataset released by NVIDIA in 2025, accompanied by the paper OpenCodeReasoning: Advancing Data Distillation for Competitive Coding, aimed at enhancing large language models' capabilities in competitive programming, code review, and SQL tasks.

The dataset contains a total of 844,935 samples, with a storage size of approximately 91 GB, in JSONL format. It covers Python, C++, programming exercises, and Text-to-SQL data, primarily used for supervised fine-tuning (SFT). The dataset is licensed under CC BY 4.0, while the Code Forces data is licensed under the Open Data Commons Attribution License (ODC-By). Other licensing information can be found in the MIT License.

Dataset Composition

• competitive_coding_python: Contains 336,568 Python competitive programming samples, covering 34,799 unique problems.
• competitive_coding_cpp: Contains 332,559 C++ competitive programming samples, used for code generation and reasoning training.
• exercism: Contains 79,244 synthetic programming exercise samples, generated based on 115 Python Exercism problems, with related problems from the Aider benchmark excluded.
• text_to_sql: Contains 96,564 synthetic SQL samples, covering dialects such as MySQL, PostgreSQL, and SQLite, with each sample including a reasoning trajectory.