README.md · HPAI-BSC/Qwen2.5-72B-Instruct-Egida-DPO at main

metadata

license: apache-2.0
datasets:
  - HPAI-BSC/Egida
language:
  - en
base_model:
  - Qwen/Qwen2.5-72B-Instruct
tags:
  - safety

Model Description

Fine-Tuned from Model: Qwen/Qwen2.5-72B-Instruct
Paper: Efficient Safety Retrofitting Against Jailbreaking for LLMs
Point of Contact: Adrián Tormos

Model Summary

This is a fine-tuned Qwen2.5-72B-Instruct model on the Egida-DPO-Qwen2.5-72B-Instruct dataset.

The Egida dataset is a collection of adversarial prompts that are thought to ellicit unsafe behaviors from language models. Specifically for this case, the Egida train split is used to run inference on Qwen2.5-72B-Instruct. Unsafe answers are selected, and paired with safe answers to create a customized DPO dataset for this model. This results in a DPO dataset composed by triplets < ”question”, ”chosen answer”, ”discarded answer” > which contain questions that elicit unsafe responses by this target model, as well as the unsafe responses produced by it.

Training Details

Hardware: NVIDIA H100 64 GB GPUs
Devices: 64 GPUs (16 nodes)
Time: 10.23h
Batch Size: 63
LR: 10−6

Performance

Safety Performance (Attack Success Ratio)

	Egida (test) ↓	DELPHI ↓	Alert-Base ↓	Alert-Adv ↓
Qwen-2.5-72B-Instruct	0.235	0.051	0.329	0.050
Qwen-2.5-72B-Instruct-Egida-DPO	0.125	0.042	0.210	0.019

General Purpose Performance

	OpenLLM Leaderboard (Average) ↑	MMLU Generative (ROUGE1) ↑
Qwen-2.5-72B-Instruct	0.618	0.771
Qwen-2.5-72B-Instruct-Egida-DPO	0.620	0.768

Refusal Ratio

	OR Bench 80K (refusal) ↓	OR Bench Hard (refusal) ↓
Qwen-2.5-72B-Instruct	0.015	0.102
Qwen-2.5-72B-Instruct-Egida-DPO	0.016	0.170

Note that this refusal ratio is computed as keyword matching with a curated list of keywords. For more information, check the paper.

Environmental Impact

Citation Information

@misc{garciagasulla2025efficientsafetyretrofittingjailbreaking,
      title={Efficient Safety Retrofitting Against Jailbreaking for LLMs}, 
      author={Dario Garcia-Gasulla and Adrian Tormos and Anna Arias-Duart and Daniel Hinjos and Oscar Molina-Sedano and Ashwin Kumar Gururajan and Maria Eugenia Cardello},
      year={2025},
      eprint={2502.13603},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.13603}, 
}