Course-Correction: Safety Alignment Using Synthetic Preferences
pillowsofwind
Abstract
The risk of harmful content generated by large language models (LLMs) becomes a critical concern. This paper presents a systematic study on assessing and improving LLMs' capability to perform the task of course-correction, \ie, the model can steer away from generating harmful content autonomously. To start with, we introduce the C^2-Eval benchmark for quantitative assessment and analyze 10 popular LLMs, revealing varying proficiency of current safety-tuned LLMs in course-correction. To improve, we propose fine-tuning LLMs with preference learning, emphasizing the preference for timely course-correction. Using an automated pipeline, we create C^2-Syn, a synthetic dataset with 750K pairwise preferences, to teach models the concept of timely course-correction through data-driven preference learning. Experiments on 2 LLMs, Llama2-Chat 7B and Qwen2 7B, show that our method effectively enhances course-correction skills without affecting general performance. Additionally, it effectively improves LLMs' safety, particularly in resisting jailbreak attacks.
Community
Our latest paper delves into LLMs' ability to perform safety self-correction, namely COURSE-CORRECTION.
In this paper, we:
- Benchmark course-correction ability
- Improving using synthetic preferences.
Paper: https://arxiv.org/pdf/2407.16637
Code: https://github.com/pillowsofwind/Course-Correction
(Figure 2) 🔰To start with, we quantitatively assess current open-source LLMs' ability to perform safety course-correction by counting the CORRECTIVE decoding paths.
(Figure 3) 📐After evaluating 10 LLMs, we found some characteristics:
- The course-correction capabilities of different SAFETY-tuned models vary widely. 😰
- For some LLMs, the more initially harmful content, the EASIER it is to perform course-correct. 😹
(Figure 4) 🏹To improve, the strategy is really simple, we craft 750K synthetic preferences by following two value principles:
- Correction is better than not
- Early correction is better than a later one
(Figure 5,6,7)We apply our synthetic data to DPO training and find that:
👉improve the course-correction ability and overall safety
👉improve the robustness of 4 jailbreak attacks
👉no harm to overall performance
👉lifted safety token probs on later decoding positions
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- SafeInfer: Context Adaptive Decoding Time Safety Alignment for Large Language Models (2024)
- Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training (2024)
- SAFETY-J: Evaluating Safety with Critique (2024)
- Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing (2024)
- Safety Arithmetic: A Framework for Test-time Safety Alignment of Language Models by Steering Parameters and Activations (2024)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment:
@librarian-bot
recommend
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper