Summary of Mix Data or Merge Models? Optimizing For Diverse Multi-task Learning, by Aakanksha et al.
Mix Data or Merge Models? Optimizing for Diverse Multi-Task Learning
by Aakanksha, Arash Ahmadian, Seraphina Goldfarb-Tarrant, Beyza Ermis, Marzieh Fadaee, Sara Hooker
First submitted to arxiv on: 14 Oct 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Machine Learning (cs.LG)
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 Large Language Models (LLMs) are widely used globally, but ensuring their safe deployment remains a significant challenge. Existing safety protocols often overfit to Western-centric datasets and fail to adapt to multilingual settings. In this work, we investigate model merging in a diverse multi-task setting, combining safety and general-purpose tasks within a multilingual context. Each language poses unique learning challenges across tasks. We find that objective-based merging outperforms data mixing by up to 8% and 10% in general performance and safety respectively. Additionally, language-based merging is highly effective, achieving a 4% increase in general performance and 7% reduction in harm across all languages on top of the data mixtures method using the same available data. Our study provides a framework for building strong and safe multilingual models. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large Language Models are used around the world, but making sure they’re used safely is a big problem. Right now, safety measures often only work well with Western languages and don’t adapt to other languages. In this research, we explore how to combine different language models in a way that makes them safer and more effective. We found that combining models based on what they’re good at (like general knowledge or safety) works better than mixing their data together. We also found that combining models from the same language can make them even stronger and safer. |
Keywords
» Artificial intelligence » Multi task
