Summary of Refine Large Language Model Fine-tuning Via Instruction Vector, by Gangwei Jiang et al.
Refine Large Language Model Fine-tuning via Instruction Vector
by Gangwei Jiang, Zhaoyi Li, Defu Lian, Ying Wei
First submitted to arxiv on: 18 Jun 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: None
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 research paper explores the phenomenon of large language models (LLMs) losing their general capabilities when fine-tuned for specific tasks. The authors identify instruction following as the primary contributor to this “forgetting” during training. To understand and mitigate this issue, they propose the Instruction Vector (IV) framework, which captures model representations related to specific instruction-following capabilities. By analyzing IV dynamics before and after training, the authors suggest that fine-tuning primarily adds specialized reasoning patterns rather than erasing previous skills. Building on this insight, they develop an IV-guided training approach aimed at preserving the original computation graph, thereby reducing catastrophic forgetting. Empirical tests on three benchmarks confirm the efficacy of this new method. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models can forget their general abilities when trained for specific tasks. Researchers studied why this happens and found that it’s mainly due to instruction following. They created a framework called Instruction Vector (IV) to understand what’s happening inside these models during training. By looking at how IVs change before and after training, they discovered that the model is actually adding new skills rather than losing old ones. To fix this problem, they developed a new way of training that preserves the original abilities of the model. This approach worked well in tests on three different datasets. |
Keywords
» Artificial intelligence » Fine tuning
