Summary of Training on the Benchmark Is Not All You Need, by Shiwen Ni et al.
Training on the Benchmark Is Not All You Need
by Shiwen Ni, Xiangtao Kong, Chengming Li, Xiping Hu, Ruifeng Xu, Jia Zhu, Min Yang
First submitted to arxiv on: 3 Sep 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
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 The proposed paper aims to develop an efficient method for detecting data leakage in pre-trained Large Language Models (LLMs) using multiple-choice benchmarks. The opacity of the pre-training process and training data can lead to unreliable results, hindering the health of the field. To address this issue, the authors suggest a simple yet effective approach based on shuffling the contents of multiple-choice options and detecting anomalies in the model’s log probability distribution over derived datasets. This method works under gray-box conditions without accessing model training data or weights. The paper demonstrates the effectiveness of the proposed method through experiments with two LLMs and benchmark designs, and evaluates the degree of data leakage in 35 mainstream open-source LLMs on four benchmark datasets, identifying Qwen family LLMs as having the highest degree of data leakage. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper wants to help make sure that large language models are not cheating by looking at test questions ahead of time. To do this, they’re proposing a new way to check if a model is doing something sneaky. They think this will be really helpful because right now, it’s hard to know for sure what’s going on when these models are tested. The method works without needing to see the actual training data or the model itself. It just looks at how well the model does on different versions of the same question and flags anything that seems suspicious. The authors tested their idea with a couple of language models and some test questions, and it seemed to work pretty well. They also looked at 35 other language models and found that some of them were really good at cheating! |
Keywords
» Artificial intelligence » Probability
