PEFT
Safetensors
query-expansion
dense-retrieval

ExpandR query expansion model

How can a language model learn which query expansions help a dense retriever rank a relevant document higher? ExpandR samples candidate expansions and uses retrieval feedback and a self-consistency signal to build preference pairs. Direct preference optimization (DPO) trains the generator on those pairs. The method section describes how the preferences are constructed.

This repository contains the LoRA adapter for the query expansion generator. The training script uses Meta-Llama-3-8B-Instruct. The original release lists PEFT 0.15.2. Use it to generate expansion text, then use the ExpandR retriever to rank documents.

Training data · Training and evaluation code

Citation

The first preprint appeared as LLM-QE: Improving Query Expansion by Aligning Large Language Models with Ranking Preferences. The archival citation is Sijia Yao, Pengcheng Huang, Zhenghao Liu, Yu Gu, Yukun Yan, Shi Yu, and Ge Yu (EMNLP 2025), ExpandR: Teaching Dense Retrievers Beyond Queries with LLM Guidance. BibTeX

Downloads last month
34
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for chengpingan/ExpandR_LLM

Adapter
(1,224)
this model

Dataset used to train chengpingan/ExpandR_LLM

Collection including chengpingan/ExpandR_LLM

Paper for chengpingan/ExpandR_LLM