Summary of Semantic Loss Guided Data Efficient Supervised Fine Tuning For Safe Responses in Llms, by Yuxiao Lu et al.
Semantic Loss Guided Data Efficient Supervised Fine Tuning for Safe Responses in LLMs
by Yuxiao Lu, Arunesh Sinha, Pradeep Varakantham
First submitted to arxiv on: 7 Dec 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed method addresses the issue of Large Language Models (LLMs) generating unsafe responses to toxic prompts by leveraging a small set of unsafe responses from the LLM itself. The approach combines a semantic cost with a negative Earth Mover Distance (EMD) loss, guiding the LLM away from generating unsafe responses. Additionally, a novel lower bound for EMD loss enables more efficient optimization. Compared to baselines, the results demonstrate superior performance and data efficiency. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large Language Models can sometimes give bad answers when asked mean questions. To fix this problem, researchers have been trying different approaches. Some methods require a lot of human help, while others rely on other language models to correct the mistakes. In this study, we propose a new method that only needs a small number of bad responses from the original model. By using a special combination of costs and losses, our approach helps the LLM avoid giving bad answers. We also introduce a new way to optimize our method efficiently. Our results show that our approach outperforms other methods in terms of both performance and data efficiency. |
Keywords
» Artificial intelligence » Optimization
