Built independently by an author, for readers. Read the story and support ChapterPal
ChapterPal Logo
Log inSign up

keyword

edit similarity

Edit similarity is a quantitative measure of how closely two sequences of characters or tokens match, determined by the minimal number of elementary edit operations, such as insertions, deletions, and substitutions, required to transform one sequence into another. Typically derived from edit distance metrics like the Levenshtein distance and often normalized against sequence lengths to produce a score between zero and one, a higher value signifies greater likeness and fewer necessary modifications between the sequences. In computational linguistics, computer science, and data processing, this metric is widely utilized for approximate string matching, spell checking, bioinformatics sequence alignment, and near-duplicate text detection to identify and group records that differ only by minor variations, typos, or slight edits.

1 item

Deduplicating Training Data Makes Language Models Better

Deduplicating Training Data Makes Language Models Better

Katherine Lee, Daphne Ippolito, Andrew Nystrom, Chiyuan Zhang, Douglas Eck, Chris Callison-Burch, Nicholas Carlini

OrganizationsGoogleUniversity of Pennsylvania

Why you should read this

Demonstrates that removing duplicate strings and documents from language model training sets drastically reduces verbatim memorization and training costs while maintaining or improving perplexity.

We find that existing language modeling datasets contain many near-duplicate examples and long repetitive substrings. As a result, over 1% of the unprompted output of language models trained on these datasets is copied verbatim from the training data. We develop two tools that allow us to deduplicate training datasets—for example removing from C4 a single 61 word English sentence that is repeated over 60,000 times. Deduplication allows us to train models that emit memorized text ten times less frequently and require fewer training steps to achieve the same or better accuracy. We can also reduce train-test overlap, which affects over 4% of the validation set of standard datasets, thus allowing for more accurate evaluation. Code for deduplication is released at https://github.com/google-research/deduplicate-text-datasets.

Added

2026-09-26