MFTCoder

Why we included this project

MFTCoder takes a different route to fine-tuning code models: instead of adapting a model to one task at a time, it trains on several related code tasks together, which the project argues is both more accurate and more efficient. The framework offers two training paths, one built on PEFT and one on accelerate, and the accelerate path supports QLoRA with DeepSpeed ZeRO-3 or FSDP, so large models can be fine-tuned without a big GPU cluster. The latest release also adds preference alignment methods (DPO, RPO, ORPO) for steering output behavior rather than just teaching new tasks. Two peer-reviewed papers, one at KDD 2024 and one at EMNLP 2024, explain the multitask and convergence-balancing methods, which helps if you want the theory behind the code, not just a script to run.