Summary of What Is Wrong with Perplexity For Long-context Language Modeling?, by Lizhe Fang et al.
What is Wrong with Perplexity for Long-context Language Modeling?
by Lizhe Fang, Yifei Wang, Zhaoyang Liu, Chenheng Zhang, Stefanie Jegelka, Jinyang Gao, Bolin Ding, Yisen Wang
First submitted to arxiv on: 31 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 This paper addresses the limitation of perplexity (PPL) as a standard evaluation metric for large language models (LLMs) in handling long-context inputs. Recent approaches have extended context windows, but PPL has proven unreliable for assessing long-context capabilities due to overlooking key tokens essential for understanding. The authors propose LongPPL, a novel metric that focuses on key tokens by employing a long-short context contrastive method, and demonstrate its strong correlation with performance on various long-context benchmarks. Additionally, they introduce LongCE loss, a re-weighting strategy for fine-tuning that prioritizes key tokens, leading to consistent improvements across diverse benchmarks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Long language models need to understand long contexts to have extended conversations, summarize documents, or learn from many examples. A problem with the way we evaluate these models is that it doesn’t account for important words and phrases. The authors of this paper explain why this is a problem and propose two new ways to measure how well models do in understanding long contexts. One method, called LongPPL, looks at important words and phrases to give a more accurate score. The other method, called LongCE loss, helps fine-tune the model to focus on these important words and phrases. |
Keywords
» Artificial intelligence » Fine tuning » Perplexity
