Improving language models by retrieving from trillions of tokens

TL;DR

This paper introduces RETRO, a retrieval-augmented language model that uses a large external database to improve performance, matching larger models with fewer parameters and enabling better downstream task results.

Contribution

The paper presents a novel retrieval-augmented transformer architecture that leverages a trillion-token database, significantly reducing model size while maintaining high performance.

Findings

RETRO achieves performance comparable to GPT-3 and Jurassic-1 with 25x fewer parameters.

Retrieval-augmented models improve downstream question answering tasks.

The approach enables rapid fine-tuning of pre-trained transformers with retrieval.

Abstract

We enhance auto-regressive language models by conditioning on document chunks retrieved from a large corpus, based on local similarity with preceding tokens. With a $2$ trillion token database, our Retrieval-Enhanced Transformer (RETRO) obtains comparable performance to GPT-3 and Jurassic-1 on the Pile, despite using 25$\times$ fewer parameters. After fine-tuning, RETRO performance translates to downstream knowledge-intensive tasks such as question answering. RETRO combines a frozen Bert retriever, a differentiable encoder and a chunked cross-attention mechanism to predict tokens based on an order of magnitude more data than what is typically consumed during training. We typically train RETRO from scratch, yet can also rapidly RETROfit pre-trained transformers with retrieval and still achieve good performance. Our work opens up new avenues for improving language models through explicit memory at unprecedented scale.