Open R1: Update #3

Hugging Face Blog · 1 year ago

Researchers released CodeForces-CoTs, a dataset of nearly 100,000 chain-of-thought samples distilled from DeepSeek-R1 for competitive programming in C++ and Python, along with OlympicCoder models fine-tuned on this data. OlympicCoder-32B outperformed Claude 3.7 Sonnet and achieved higher scores than o1-mini and DeepSeek-R1 on the IOI 2024 benchmark, despite being 7 billion to 32 billion parameters. The release enables open-source development of code reasoning models and provides a new fully-verifiable benchmark from International Olympiad in Informatics problems, addressing limitations in existing competitive programming datasets that only include partial test cases.