Balanced Data Sampling for Language Model Training with Clustering

TL;DR

This paper introduces ClusterClip Sampling, a clustering-based data sampling method for training large language models that balances data distribution and reduces overfitting, leading to improved performance over traditional random sampling.

Contribution

The paper proposes a novel clustering-based sampling strategy, ClusterClip Sampling, to balance training data and mitigate overfitting in large language model training.

Findings

ClusterClip Sampling outperforms random sampling across various datasets.

It effectively balances common and rare samples during training.

The method reduces overfitting caused by overrepresented clusters.

Abstract

Data plays a fundamental role in the training of Large Language Models (LLMs). While attention has been paid to the collection and composition of datasets, determining the data sampling strategy in training remains an open question. Most LLMs are trained with a simple strategy, random sampling. However, this sampling strategy ignores the unbalanced nature of training data distribution, which can be sub-optimal. In this paper, we propose ClusterClip Sampling to balance the text distribution of training data for better model training. Specifically, ClusterClip Sampling utilizes data clustering to reflect the data distribution of the training set and balances the common samples and rare samples during training based on the cluster results. A repetition clip operation is introduced to mitigate the overfitting issue led by samples from certain clusters. Extensive experiments validate the effectiveness of ClusterClip Sampling, which outperforms random sampling and other cluster-based sampling variants under various training datasets and large language models.