Reproduce “Attributing Culture-Conditioned Generations to Pretraining Corpora” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Huihan Li, Arnav Goel, Keyu He, Xiang Ren
The authors introduce MEMOed, a method for tracing culture-conditioned food and clothing outputs to relevant pretraining documents. It compares how strongly documents link a culture and generated symbol across cultures. In OLMo-7B, 97/110 food cultures and 45/110 clothing cultures had at least one memorized symbol; results are limited to this model and searchable Dolma data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.