MFTCoder is an open-source multi-task fine-tuning framework for code large language models, developed by Ant Group’s CodeFuse team. It is designed to raise both the accuracy and the training efficiency of code LLMs across many tasks at once. The work was accepted at KDD 2024, and the repository has about 711 stars on GitHub.
Code models are usually fine-tuned separately for completion, generation, translation, repair and other tasks, which wastes data and compute. MFTCoder instead trains one model on many code tasks jointly, with techniques that balance the tasks so none dominates or degrades the others. The framework supports multiple base models and is built to be customizable, letting researchers plug in their own task mixtures and configurations. Multi-task learning also tends to improve generalization, so a model trained this way handles unseen coding requests more robustly than single-task specialists. The repository was last updated in December 2024 and remains a reference implementation for multi-task code model training.
In practice, that means a team can maintain one strong all-round code model instead of a zoo of task-specific checkpoints, which simplifies deployment, evaluation and ongoing maintenance of an internal coding assistant.
MFTCoder is free and open source. You can download the framework from GitHub and use it with your own base models and training infrastructure; the only costs are your compute and data.
MFTCoder is for ML engineers and researchers who train or fine-tune code language models and want one model that performs well across completion, generation and repair tasks. It is particularly useful to teams maintaining internal code assistants on open base models.