Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2024
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2501.06208 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866911102607032320 |
|---|---|
| author | Gudipudi, Satya Swaroop Vipparla, Sreeram Singh, Harpreet Goel, Shashwat Kumaraguru, Ponnurangam |
| author_facet | Gudipudi, Satya Swaroop Vipparla, Sreeram Singh, Harpreet Goel, Shashwat Kumaraguru, Ponnurangam |
| contents | Instruction fine-tuning of large language models (LLMs) is a powerful method for improving task-specific performance, but it can inadvertently lead to a phenomenon where models generate harmful responses when faced with malicious prompts. In this paper, we explore Low-Rank Adapter Fusion (LoRA) as a means to mitigate these risks while preserving the model's ability to handle diverse instructions effectively. Through an extensive comparative analysis against established baselines using recognized benchmark datasets, we demonstrate a 42\% reduction in the harmfulness rate by leveraging LoRA fusion between a task adapter and a safety adapter, the latter of which is specifically trained on our safety dataset. However, we also observe exaggerated safety behaviour, where the model rejects safe prompts that closely resemble unsafe ones |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2501_06208 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Enhancing AI Safety Through the Fusion of Low Rank Adapters Gudipudi, Satya Swaroop Vipparla, Sreeram Singh, Harpreet Goel, Shashwat Kumaraguru, Ponnurangam Computation and Language Instruction fine-tuning of large language models (LLMs) is a powerful method for improving task-specific performance, but it can inadvertently lead to a phenomenon where models generate harmful responses when faced with malicious prompts. In this paper, we explore Low-Rank Adapter Fusion (LoRA) as a means to mitigate these risks while preserving the model's ability to handle diverse instructions effectively. Through an extensive comparative analysis against established baselines using recognized benchmark datasets, we demonstrate a 42\% reduction in the harmfulness rate by leveraging LoRA fusion between a task adapter and a safety adapter, the latter of which is specifically trained on our safety dataset. However, we also observe exaggerated safety behaviour, where the model rejects safe prompts that closely resemble unsafe ones |
| title | Enhancing AI Safety Through the Fusion of Low Rank Adapters |
| topic | Computation and Language |
| url | https://arxiv.org/abs/2501.06208 |