Scalable Multi-Stage Influence Function for Large Language Models via Eigenvalue-Corrected Kronecker-Factored Parameterization

TL;DR

This paper introduces a scalable multi-stage influence function for large language models, enabling attribution of predictions to pre-training data with improved efficiency using EK-FAC parameterization, validated through empirical experiments and case studies.

Contribution

It proposes a novel multi-stage influence function for fine-tuned LLMs and leverages EK-FAC for scalable approximation, addressing previous limitations in interpretability and scalability.

Findings

EK-FAC provides superior scalability for influence computation.

Multi-stage influence functions effectively attribute predictions to training data.

Case studies demonstrate interpretive insights for large LLMs.

Abstract

Pre-trained large language models (LLMs) are commonly fine-tuned to adapt to downstream tasks. Since the majority of knowledge is acquired during pre-training, attributing the predictions of fine-tuned LLMs to their pre-training data may provide valuable insights. Influence functions have been proposed as a means to explain model predictions based on training data. However, existing approaches fail to compute ``multi-stage'' influence and lack scalability to billion-scale LLMs. In this paper, we propose the multi-stage influence function to attribute the downstream predictions of fine-tuned LLMs to pre-training data under the full-parameter fine-tuning paradigm. To enhance the efficiency and practicality of our multi-stage influence function, we leverage Eigenvalue-corrected Kronecker-Factored (EK-FAC) parameterization for efficient approximation. Empirical results validate the superior scalability of EK-FAC approximation and the effectiveness of our multi-stage influence function. Additionally, case studies on a real-world LLM, dolly-v2-3b, demonstrate its interpretive power, with exemplars illustrating insights provided by multi-stage influence estimates. Our code is public at https://github.com/colored-dye/multi_stage_influence_function.