# InPars-v2: Large Language Models as Efficient Dataset Generators for Information Retrieval

**Authors:** Vitor Jeronymo, Luiz Bonifacio, Hugo Abonizio, Marzieh Fadaee, Roberto Lotufo, Jakub Zavrel, Rodrigo Nogueira  
**arXiv:** [2301.01820](https://arxiv.org/abs/2301.01820)  
**Date:** 2023-01-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2301.01820v1/

## Key points

- InPars-v2 is a novel approach to information retrieval that leverages large language models (LLMs) to generate synthetic query-document pairs.
- It builds upon the previous work of Bonifacio et al. in InPars, enhancing it by incorporating a reranker as a filtering mechanism for identifying relevant synthetic examples.
- The key distinction between InPars and Promptagator is in their implementations: InPars uses proprietary LLMs like GPT-3 and FLAN for dataset generation, while Promptagator employs dataset-specific prompts and a larger LLM alongside a fully trainable retrieval pipeline with smaller models.
- Inpars-v1 and Inpars-v2 represent iterative advancements in leveraging LLMs for dataset generation in information retrieval tasks.
- The authors opt for an open-source query generator over proprietary tools used in prior research, making their methodology more accessible and reproducible.
- Detailed insights into the experimental setup highlight the use of open-source GPT-J for synthetic query generation and emphasize reproducibility through shared source code and data tailored for TPUs.

## Summary

InPars-v2: A Novel Approach to Information Retrieval Using LLMs and Synthetic Query-Document Pairs

Bonifacio et al. have introduced a novel approach, InPars-v2, which builds upon their previous work InPars. This latest iteration leverages large language models (LLMs) to generate synthetic query-document pairs for information retrieval tasks. Similar to Promptagator, another recent model that utilizes LLMs for creating alternative queries in an unsupervised manner. However, the key distinction between InPars and Promptagator lies in their implementations. While InPars uses proprietary LLMs like GPT-3 and FLAN for dataset generation, Promptagator employs dataset-specific prompts and a larger LLM alongside a fully trainable retrieval pipeline with smaller models. Bonifacio et al. enhance the original InPars framework by incorporating a reranker as a filtering mechanism to identify the most relevant synthetic examples. This enhancement results in superior performance on the BEIR benchmark. Furthermore, the authors opt for an open-source query generator over proprietary tools used in prior research, making their methodology more accessible and reproducible. Referred to as Inpars-v1 and Inpars-v2 respectively, these models represent iterative advancements in leveraging LLMs for dataset generation in information retrieval tasks. The authors provide detailed insights into their experimental setup, highlighting the use of open-source GPT-J for synthetic query generation and emphasizing the reproducibility of their results through shared source code and data tailored for TPUs. Overall, this refined summary showcases how Bonifacio et al. 's innovative approach not only builds upon existing research but also sets new benchmarks in information retrieval through efficient dataset generation using open-source tools and advanced reranking techniques.