Summary of Selecting Large Language Model to Fine-tune Via Rectified Scaling Law, by Haowei Lin et al.
Selecting Large Language Model to Fine-tune via Rectified Scaling Law
by Haowei Lin, Baizhou Huang, Haotian Ye, Qinyu Chen, Zihao Wang, Sujian Li, Jianzhu Ma, Xiaojun Wan, James Zou, Yitao Liang
First submitted to arxiv on: 4 Feb 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 research paper tackles the challenge of selecting the most suitable pre-trained language model to fine-tune in a resource-constrained environment. The authors formulate this task as predicting fine-tuning performance and demonstrate its connection with Scaling Law, which has been widely used in deep learning. The study reveals that the fine-tuning scaling curve exhibits two phases: the well-known “power phase” and a previously unobserved “pre-power phase”. They also show why existing Scaling Law fails to capture this phenomenon both theoretically and empirically. To address this limitation, the authors introduce the concept of “pre-learned data size” into their Rectified Scaling Law, which better fits experimental results and provides a novel LLM selection algorithm. This algorithm selects the near-optimal model with significantly reduced resource consumption. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper solves a big problem in machine learning. Imagine you have many pre-trained models to choose from, but you don’t have enough resources to test them all. The authors came up with a clever way to predict how well each model will perform when fine-tuned. They discovered that the process of fine-tuning has two stages: one where performance improves quickly and another where it slows down. This is important because existing methods didn’t capture this pattern. To fix this, they developed a new “Rectified Scaling Law” that helps us choose the best model with much less effort. |
Keywords
* Artificial intelligence * Deep learning * Fine tuning * Language model * Machine learning
