Restoring Calibration for Aligned Large Language Models

ICML 2025 | Paper: Restoring Calibration for Aligned Large Language Models: A Calibration-Aware Fine-Tuning Approach

📄 Paper on OpenReview

---

🧰 Code Structure

.
├── scripts/             # Training & evaluation scripts
├── models/              # LoRA + Fine-tuning code
├── calibrate/           # CFT and RCFT methods
├── data/                # Dataset processing
├── plots/               # Calibration visualizations
└── README.md            # This file

---

🚀 Getting Started

1. Environment Setup

conda create -n llm-calibration python=3.10
conda activate llm-calibration
pip install -r requirements.txt

---

📌 Citation

@inproceedings{xiao2025restoring,
  title     = {Restoring Calibration for Aligned Large Language Models: A Calibration-Aware Fine-Tuning Approach},
  author    = {Xiao, Jiancong and Hou, Bojian and Wang, Zhanliang and Jin, Ruochen and Long, Qi and Su, Weijie J. and Shen, Li},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2025}
}

---

🤝 Acknowledgements

This project is developed by researchers at the University of Pennsylvania. We thank the open-source community and prior foundational work on LLM calibration, DPO, and RLHF.