Open R1: Update #3

Over the last few weeks, we have focused our efforts on reproducing the competitive programming (code reasoning) aspects of the DeepSeek-R1 recipe.

Here’s an overview of how the OlympicCoder models stack up against various instruction fine-tuned and reasoning models. We find that training models on CodeForces-CoTs produces top-tier performance, with OlympicCoder-32B outperforming all open-weight models we tested, including some that are over 100x larger 🤯.

Read on to learn how we built the dataset, benchmark and models!

In this solution, we define a function add_numbers that takes two parameters a and b and returns their sum. We then call this function with 1 and 1, storing the result in the variable result , which is then printed. This demonstrates a straightforward way to compute and display the sum of two integers. The example serves as a simple yet effective illustration of basic arithmetic operations in Python.

The problems are extremely challenging, and the full test sets are available and released under a permissive (CC-BY) license. This means that IOI is the perfect dataset to test a model's code reasoning capabilities.

CodeForces is one of the most popular websites among competitive programmers, hosting regular contests where participants must solve challenging algorithmic optimization problems. The challenging nature of these problems makes them an interesting dataset to improve and test models’ code reasoning capabilities.

While previous efforts such as DeepMind’s CodeContests dataset have compiled a large number of CodeForces problems, today we are releasing our own open-r1/codeforces dataset, with more than 10k problems covering the very first contests all the way to 2025, ~3k of which were not included in DeepMind’s dataset. Additionally, for around 60% of the problems, we have included the editorial , which is an explanation, written by the contest organizers, explaining the correct solution. You will also find 3 correct solutions per problem extracted from the official website.

Furthermore, we are releasing open-r1/codeforces-cots , which contains chain of thought generations produced by DeepSeek-R1 on these problems, where we asked the model to produce solutions in C++ (the main language used in competitive programming) and Python, totaling close to 100k samples.