Reproduce “Token-wise Influential Training Data Retrieval for Large Language Models” 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

Huawei Lin, Jikai Long, Zhaozhuo Xu, Weijie Zhao

RapidIn introduces a scalable and efficient framework for token-wise influential training data retrieval in Large Language Models (LLMs). It overcomes the computational and memory limitations of prior influence estimation methods, making it practical for models with billions of parameters and trillions of tokens.

Key insights

  • Gradient compression, achieved through efficient random shuffling and random projection, can preserve sufficient information for accurate influence estimation while drastically reducing memory footprint.
  • Layer-wise normalization of gradients stabilizes their magnitudes across deep networks, making influence calculations more robust and preventing disproportionate weighting from certain layers.
  • Decomposing influence calculations to the token level provides fine-grained interpretability, enabling insights into which specific tokens in training data influenced particular tokens in model output.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.