the official repository for the paper "COFFEE: Boost Your CodeLLMs by Fixing Bugs with Feedback".
Overview of framework, COFFEEPOTs (COde Fixing with FEEdback via Preference-Optimized Tuning and Selection)
We are excited to share our dataset and model checkpoints with the community!
You can now access them via Hugging Face:
- COFFEE Dataset
A curated dataset for feedback-driven editing.
- Coffee-Critic – Baseline critic model trained for feedback generation
- Coffee-Editor – Baseline Editor model trained for code generation
- Coffee-DPO – Critic model optimized through Direct Preference Optimization (DPO)
- Coffee-Selector – Selector model for choosing the most appropriate feedback
For any questions about the implementation or content of the paper, you could contact us via the following email :)
lune-blue@yonsei.ac.kr
kopf_yhs@yonsei.ac.kr
mapoout@yonsei.ac.kr