CWRUSafetyLab/Qwen2.5-1.5B-Instruct-EASE
CWRUSafetyLab/Qwen2.5-1.5B-Instruct-EASE is a fine-tuned Qwen2.5-1.5B-Instruct model developed by CWRUSafetyLab. This model is specifically aligned for adaptive safety reasoning, activating explicit safety measures only under jailbreak-like prompts. It aims to maintain general task effectiveness and efficiency while enhancing robustness against jailbreak attacks. This makes it suitable for safety-oriented research in small language models and jailbreak robustness.
Loading preview...
Overview
CWRUSafetyLab/Qwen2.5-1.5B-Instruct-EASE is a specialized instruction-tuned language model based on Qwen2.5-1.5B-Instruct. Developed by CWRUSafetyLab, this model is fine-tuned on the EASE-SafetyReasoning dataset, which is part of the EASE framework for practical and efficient safety alignment in small language models.
Key Capabilities
- Adaptive Safety Reasoning: The model is designed to activate explicit safety reasoning only when it detects jailbreak-like semantics in prompts.
- Efficiency and Effectiveness: It avoids unnecessary safety reasoning on benign or general prompts, thereby preserving the model's general task performance and computational efficiency.
- Jailbreak Robustness: A primary goal of this fine-tuning is to improve the model's resilience against various jailbreak attacks.
Intended Use Cases
This model is primarily intended for safety-oriented research, focusing on:
- Safety Alignment: Investigating and developing methods for aligning language models with safety principles.
- Small Language Models (SLMs): Researching the unique challenges and opportunities in safety for smaller models.
- Jailbreak Robustness: Studying and enhancing the ability of models to resist malicious prompts designed to bypass safety filters.
For more details, refer to the associated paper: EASE: Practical and Efficient Safety Alignment for Small Language Models.