Scaling Retrieval-Based Language Models with a Trillion-Token Datastore

TL;DR

This paper demonstrates that increasing the size of the datastore in retrieval-based language models consistently enhances performance across tasks, with a 1.4 trillion-token datastore outperforming traditional models, highlighting datastore scaling as a key factor.

Contribution

The authors introduce the largest open-source 1.4 trillion-token datastore and provide a comprehensive analysis of datastore scaling effects on language model performance.

Findings

Larger datastores improve language modeling and downstream tasks monotonically.

A smaller model with a large datastore outperforms bigger models without retrieval.

Datastore size significantly enhances performance for the same compute budget.

Abstract

Scaling laws with respect to the amount of training data and the number of parameters allow us to predict the cost-benefit trade-offs of pretraining language models (LMs) in different configurations. In this paper, we consider another dimension of scaling: the amount of data available at inference time. Specifically, we find that increasing the size of the datastore used by a retrieval-based LM monotonically improves language modeling and several downstream tasks without obvious saturation, such that a smaller model augmented with a large datastore outperforms a larger LM-only model on knowledge-intensive tasks. By plotting compute-optimal scaling curves with varied datastore, model, and pretraining data sizes, we show that using larger datastores can significantly improve model performance for the same training compute budget. We carry out our study by constructing a 1.4 trillion-token datastore named MassiveDS, which is the largest and the most diverse open-sourced datastore for retrieval-based LMs to date, and designing an efficient pipeline for studying datastore scaling in a computationally accessible manner. Finally, we analyze the effect of improving the retriever, datastore quality filtering, and other design choices on our observed scaling trends. Overall, our results show that datastore size should be considered as an integral part of LM efficiency and performance trade-offs. To facilitate future research, we open-source our datastore and code at https://github.com/RulinShao/retrieval-scaling.