Reproduce “CHEF: A Cheap and Fast Pipeline for Iteratively Cleaning Label Uncertainties (Technical Report)” 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
Yinjun Wu, James Weimer, Susan B. Davidson
CHEF addresses weak-label repair by ranking uncertain examples and suggesting replacement labels according to estimated effects on trusted validation loss. It processes examples in small batches, reusing computations to update the model between rounds. On Twitter sentiment, after cleaning 100 samples in batches of 10, the INFL-suggested-label strategy reaches held-out F1 0.6697 ± 0.0058, versus 0.6485 without cleaning. This result applies to this task and setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.