Self-Influence Guided Data Reweighting for Language Model Pre-training

TL;DR

This paper introduces PRESENCE, a novel method for reweighting pre-training data in language models using self-influence scores, aiming to improve training efficiency and model quality by prioritizing more relevant and high-quality samples.

Contribution

The paper proposes PRESENCE, the first model-driven data reweighting approach for language model pre-training that leverages self-influence scores to enhance training stability and effectiveness.

Findings

PRESENCE improves model performance across multiple datasets.

Reweighting with self-influence scores enhances training stability.

The method is effective across different model sizes and tasks.

Abstract

Language Models (LMs) pre-trained with self-supervision on large text corpora have become the default starting point for developing models for various NLP tasks. Once the pre-training corpus has been assembled, all data samples in the corpus are treated with equal importance during LM pre-training. However, due to varying levels of relevance and quality of data, equal importance to all the data samples may not be the optimal choice. While data reweighting has been explored in the context of task-specific supervised learning and LM fine-tuning, model-driven reweighting for pre-training data has not been explored. We fill this important gap and propose PRESENCE, a method for jointly reweighting samples by leveraging self-influence (SI) scores as an indicator of sample importance and pre-training. PRESENCE promotes novelty and stability for model pre-training. Through extensive analysis spanning multiple model sizes, datasets, and tasks, we present PRESENCE as an important first step in the research direction of sample reweighting for pre-training language models.