Summary of Adaptive Pre-training Data Detection For Large Language Models Via Surprising Tokens, by Anqi Zhang et al.
Adaptive Pre-training Data Detection for Large Language Models via Surprising Tokens
by Anqi Zhang, Chaofeng Wu
First submitted to arxiv on: 30 Jul 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Cryptography and Security (cs.CR); 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 research paper proposes an adaptive method for detecting pre-training data used in large language models (LLMs). The current solutions rely on the memorization capabilities of LLMs, but this approach has limitations. Our proposed method identifies “surprising tokens” by analyzing the prediction probability and entropy of the model’s output. We show that this method can be applied without accessing the pre-training data or requiring additional training. Experimental results demonstrate a consistent improvement over existing methods on various benchmarks and models, with a maximum increase of 29.5%. Additionally, we introduce a new benchmark, Dolma-Book, which evaluates model performance using book data collected before and after training. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper tries to solve a problem where big language models are secretly trained on private data without permission. The current solutions rely on the model’s ability to memorize words, but this isn’t effective because there is too much data and not enough time to train. Our new method detects when the model has seen private data by analyzing its predictions. We show that our approach works better than existing methods in many experiments. We also create a new test for evaluating these models. |
Keywords
* Artificial intelligence * Probability
