HPAI-BSC/Qwen2.5-7B-Instruct-Egida-DPO

Model Description

• Fine-Tuned from Model: Qwen/Qwen2.5-7B-Instruct
• Paper: Efficient Safety Retrofitting Against Jailbreaking for LLMs
• Point of Contact: Adrián Tormos

Model Summary

This is a fine-tuned Qwen2.5-7B-Instruct model on the Egida-DPO-Qwen2.5-7B-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 Qwen2.5-7B-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: 4 GPUs (1 node)
• Time: 1.59h
• Batch Size: 8
• LR: 10−7

Performance

Safety Performance (Attack Success Ratio)

	Egida (test) ↓	DELPHI ↓	Alert-Base ↓	Alert-Adv ↓
Qwen-2.5-7B-Instruct	0.471	0.138	0.544	0.080
Qwen-2.5-7B-Instruct-Egida-DPO	0.322	0.118	0.410	0.045

General Purpose Performance

	OpenLLM Leaderboard (Average) ↑	MMLU Generative (ROUGE1) ↑
Qwen-2.5-7B-Instruct	0.488	0.331
Qwen-2.5-7B-Instruct-Egida-DPO	0.488	0.296

Refusal Ratio

	OR Bench 80K (refusal) ↓	OR Bench Hard (refusal) ↓
Qwen-2.5-7B-Instruct	0.021	0.175
Qwen-2.5-7B-Instruct-Egida-DPO	0.029	0.240

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}, 
}