Training data attribution seeks to identify which training examples are most related to a model prediction or user query, yet applying attribution methods to modern foundation models practice often brings difficulties to users as such methods could be computationally expensive, method-specific, and hard to inspect within a single analysis process. This thesis presents an interactive framework for training data attribution that combines live query computation with precomputed training-side reference representations, enabling users to retrieve, compare, and interpret ranked training examples for live queries or selected validation examples within a unified workflow. The framework centers on gradient-based attribution methods, including gradient similarity, DataInf-style influence approximation over projected gradient features, LESS-style low-rank gradient matching, and LoGra-style features. It also supports multi-query analysis through score aggregation and comparative inspection across methods, allowing attribution to be analyzed as a workflow rather than as an isolated score. Demonstrations are conducted on public question-answering datasets with Pythia-family models, on a multimodal driving-video setting with Qwen3-VL, and on a medical-domain setting that extends the framework to private and domain-specific data. This thesis shows that training data attribution becomes substantially more useful when it is treated not only as a scoring method, but as a practical workflow for querying, comparing, and interpreting training examples across models, methods, and datasets.

61 pages

Thesis Committee:
Chenyan Xiong (Chair)
Alexander Rudnicky

Jignesh Patel, Interim Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu