Retrieval augmentation enables large language models to take advantage of external knowledge, for example on tasks like question answering and data imputation. However, the performance of such retrieval-augmented models is limited by the data quality of their underlying retrieval corpus. In this paper, we propose an algorithm based on multilinear extension for evaluating the data importance of retrieved data points. There are exponentially many terms in the multilinear extension, and one key contribution of this paper is a polynomial time algorithm that computes exactly, given a retrieval-augmented model with an additive utility function and a validation set, the data importance of data points in the retrieval corpus using the multilinear extension of the model's utility function. We further proposed an even more efficient ({\epsilon}, {\delta})-approximation algorithm. Our experimental results illustrate that we can enhance the performance of large language models by only pruning or reweighting the retrieval corpus, without requiring further training. For some tasks, this even allows a small model (e.g., GPT-JT), augmented with a search engine API, to outperform GPT-3.5 (without retrieval augmentation). Moreover, we show that weights based on multilinear extension can be computed efficiently in practice (e.g., in less than ten minutes for a corpus with 100 million elements).

After writing this, someone asked me if there's a way to avoid to constantly pay OpenAI API and still use at scale any GPT-3-based NLP pipeline, trying to keep the same quality. Of course, if you have enough budget for your use case, a fine-tuned GPT-3 (which now turned to be the even more powerful GPT-3.5) may still be the best choice in terms of quality. Said that, I think there are strategies to use GPT-3 more wisely, save some money and still get good results. For ABSA, sentiment analysis or any other NLP text classification task, you may consider to redesign your pipeline leveraging few-shot text classification. Indeed, recently significant improvements have been made public on this kind of tasks.

Recently, Together, an open source community led by researchers and technocrats, released a new version of GPT-JT trained on six billion parameters. This new model has been built using recently published open source techniques and datasets. It has been trained with its decentralised approach on the Together Research Computer, a local, in-house network developed by the team. Unlike GPT-3, which is available on-demand/request, GPT-JT is now available as open source. Click here to access the code and datasets.