Reproduce “Fast Training Dataset Attribution via In-Context Learning” 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
Milad Fotouhi, Mohammad Taha Bahadori, Oluwaseyi Feyisetan, Payman Arabshahi, David Heckerman
The paper introduces two methods, Shapley Context Method (SCM) and Context Mixture Factorization (CMF), for attributing training dataset contributions to black-box instruction-tuned Large Language Models. CMF, in particular, offers a computationally efficient and robust approach that quantifies dataset influence, validated through fine-tuning and machine unlearning experiments.
Key insights
- If an LLM has already learned information from a dataset, providing that dataset as external context should not significantly alter its output.
- Modeling the LLM's behavior as a mixture distribution allows for identifying specific dataset contributions through observed output probabilities with and without context.
- Prompt engineering combined with RAG enables an efficient, black-box method for dataset attribution that bypasses the need for model retraining or gradient access.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.