Summary of Seal: Safety-enhanced Aligned Llm Fine-tuning Via Bilevel Data Selection, by Han Shen and Pin-yu Chen and Payel Das and Tianyi Chen
SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selection
by Han Shen, Pin-Yu Chen, Payel Das, Tianyi Chen
First submitted to arxiv on: 9 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
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 This paper proposes SEAL, a novel framework to enhance safety in Large Language Model (LLM) fine-tuning. The authors show that previous approaches can compromise the model’s pre-equipped alignment and safety capabilities by fine-tuning on adversarial or low-quality data. Instead, SEAL uses bilevel optimization to learn a data ranker that prioritizes high-quality training data. As a result, models trained with SEAL outperform baselines in terms of quality, with an 8.5% and 9.7% win rate increase for the Llama-3-8b-Instruct and Merlinite-7b models, respectively. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Fine-tuning Large Language Models to improve their performance is important, but previous studies have shown that this process can actually make the models worse if they’re trained on low-quality data. The authors of this paper propose a new way to fine-tune these models called SEAL. SEAL helps keep the models safe and effective by prioritizing high-quality training data. This results in better-performing models than other approaches, with an 8.5% and 9.7% increase in win rate for two specific models. |
Keywords
» Artificial intelligence » Alignment » Fine tuning » Large language model » Llama » Optimization
