OpenCodeReasoning
a dataset for training LLMs in coding reasoning
1 min read
NVIDIA's OpenCodeReasoning is currently the largest synthetic dataset for supervised fine-tuning (SFT) of language models in competitive programming. The collection includes over 735,000 Python samples spanning 28,319 unique problems from popular platforms like CodeForces, LeetCode, and AtCoder. Both the solutions and reasoning traces were generated using the R1 model. This dataset is ideal for developers training or fine-tuning LLMs to generate reliable, logically sound code.
Key features
- Massive scale: 735,255 Python samples across 28,319 problems.
- Synthetic responses: all solutions and reasoning chains are generated by R1.
- Diverse sources: problems are sourced from CodeForces, LeetCode, HackerRank, AtCoder, and other platforms.
- Permissive license: ready for commercial and non-commercial use under CC BY 4.0.
What you can do
- Train or fine-tune LLMs to improve their reasoning skills when writing code.
- Conduct research on knowledge distillation for software generation.
- Integrate the data into SFT pipelines using tools like NVIDIA NeMo-Skills.
Example
Loading the dataset with the datasets library:
from datasets import load_dataset
ocr_ds_split_0 = load_dataset("nvidia/OpenCodeReasoning", "split_0")
print(ocr_ds_split_0)Read more
DetailsTotal size: 735,255 samples (567,850 in split_0 and 167,405 in split_1)
Programming language: Python
Problem platforms: CodeForces, CodeChef, LeetCode, AtCoder, HackerRank, etc.
License: CC BY 4.0
Developer: NVIDIA








Comments