Summary of Improving Pretraining Data Using Perplexity Correlations, by Tristan Thrush et al.
Improving Pretraining Data Using Perplexity Correlations
by Tristan Thrush, Christopher Potts, Tatsunori Hashimoto
First submitted to arxiv on: 9 Sep 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Machine Learning (cs.LG); Machine Learning (stat.ML)
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 presents a novel framework for selecting high-quality pretraining data without requiring costly LLM training runs. The approach is based on observing the correlation between LLM losses and downstream benchmark performance, allowing for effective pretraining data selection. The authors develop a statistical framework centered around estimates of perplexity-benchmark correlations and apply it to select pretraining texts from tens of thousands of web domains using a sample of 90 LLLMs from the Open LLM Leaderboard. In controlled pretraining experiments at the 160M parameter scale on 8 benchmarks, the proposed method outperforms DSIR on every benchmark while matching the best data selector found in DataComp-LM, a hand-engineered bigram classifier. The paper also includes results from preregistered experiments with new pretraining data on an aggregation of 22 benchmarks up to the 1.4B scale, demonstrating increasing improvements of the method with more scale. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research helps create better language models by selecting the best texts to use during training. Instead of doing many expensive tests to find the right texts, this approach uses a simple idea: good texts for training are often related to how well the model performs in real-world tasks. The team developed a new way to estimate these relationships and applied it to select thousands of web pages from a large dataset. They then tested their method on several benchmarks and found that it performed better than other methods, even with smaller models. This breakthrough could lead to more powerful language models that can help us communicate more effectively. |
Keywords
» Artificial intelligence » Perplexity » Pretraining
