OpenCodeInterpreter

Why we included this project

OpenCodeInterpreter is a family of open-source code models that actually run the code they generate. Instead of stopping after producing a snippet, the models execute it, read the runtime output, and use that feedback to refine the answer; the README's benchmark tables show the gains on HumanEval and MBPP. The repo also includes the Code-Feedback dataset, the multi-turn interactions used to train the models, plus a locally deployable demo where you can chat with the model and watch it correct its own output. Teams experimenting with self-hosted coding assistants, or researchers studying how execution feedback improves generation, will find both the weights and the training data here.