language:
- en
pretty_name: DIVERGE
tags:
- rag
- retrieval-augmented-generation
- open-information-seeking
- diversity
license: mit
task_categories:
- text-generation
- question-answering
Diverge
Diverge is derived from infinite-chats. Diverge is constructed to support research on diversity-aware retrieval-augmented generation (RAG) in open information-seeking settings.
Source Data
Diverge is built upon Dataset Infinite-Chats, To focus on open-ended and opinion-diverse tasks, we define a predefined set of ten high-level categories: Problem Solving, Decision Support, Concept Explanations, Skill Development, Recommendations, Opinion-Based Questions, Value-Laden Questions, Controversial Questions, Ideation and Brainstorming, and Personal Advice.
For each conversation, we extract the user prompt by selecting the first message whose role is labeled as user. All other conversational context is discarded. We randomly iterate through the dataset sequentially and collect instances satisfying the category constraint until reaching a fixed budget.
📄 Associated Paper:
DIVERGE: Diversity-Enhanced Retrieval-Augmented Generation for Open-Ended Questions
Citation
If you use this dataset, please cite:
@article{hu2026diverge,
title={DIVERGE: Diversity-Enhanced RAG for Open-Ended Information Seeking},
author={Hu, Tianyi and Tandon, Niket and Arora, Akhil},
journal={arXiv preprint arXiv:2602.00238},
year={2026}
}