Summary of Toward the Evaluation Of Large Language Models Considering Score Variance Across Instruction Templates, by Yusuke Sakai et al.
Toward the Evaluation of Large Language Models Considering Score Variance across Instruction Templates
by Yusuke Sakai, Adam Nohejl, Jiangnan Hang, Hidetaka Kamigaito, Taro Watanabe
First submitted to arxiv on: 22 Aug 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); 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 study aims to address the unfair evaluation methods used for natural language understanding (NLU) performance of large language models (LLMs). Existing evaluation methods do not account for score variation due to differences in prompts, leading to biased comparison. The proposed solution is a fair evaluation method that considers variance between instruction templates, along with a new metric called Sharpe score. To achieve this, the study provides cross-lingual datasets for English and Japanese LLMs, including multiple instruction templates for fair evaluation of each task. Comprehensive analysis reveals that variance among templates has a significant impact on evaluating LLMs. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about how to make sure big language models are judged fairly when they’re tested on different tasks. Right now, the way we test them doesn’t take into account the fact that different prompts can get very different results. This means some models might look better than others just because of the prompt they got, not because they’re actually good at understanding language. The study solves this problem by creating special datasets and a new way to measure how well the models are doing. |
Keywords
* Artificial intelligence * Language understanding * Prompt
