We develop a methodology for analyzing language model task performance at the individual example level based on training data density estimation. Experiments with paraphrasing as a controlled intervention on finetuning data demonstrate that increasing the support in the training distribution for specific test queries results in a measurable increase in density, which is also a significant predictor of the performance increase caused by the intervention. Experiments with pretraining data demonstrate that we can explain a significant fraction of the variance in model perplexity via density measurements. We conclude that our framework can provide statistical evidence of the dependence of a target model's predictions on subsets of its training data, and can more generally be used to characterize the support (or lack thereof) in the training data for a given test task.
@article{arxiv.2405.06331,
title = {LMD3: Language Model Data Density Dependence},
author = {John Kirchenbauer and Garrett Honke and Gowthami Somepalli and Jonas Geiping and Daphne Ippolito and Katherine Lee and Tom Goldstein and David Andre},
journal= {arXiv preprint arXiv:2405.06331},
year = {2024}
}