PolyBench

PolyBench is a comprehensive benchmark to facilitate the evaluation of relevant methods on diverse question patterns for augmenting large language models with external knowledge graphs.

Dataset Details

• Curated by: Renjie Liu (https://liu-rj.github.io/)
• Language: English
• License: apache-2.0

Dataset Description

PolyBench is created on three real-world knowledge graphs from GRBench across the academia, literature and e-commerce domain. With a total of 1,200 questions instantiated from 73 well-crafted question templates, PolyBench consists of 400 questions for each graph spanning 5 question patterns. This comprehensive set of question patterns is derived from a complete four-class taxonomy under the triplet format of the knowledge graph, including four basic patterns and nested types:

• <s,*,*> (subject centered): Questions about an entity (the subject) with no specific relation constraints (the predicate) and target entity (the object). The task is to answer a general question about the entity. Example: “Who is Isaac Newton?”
• <s,p,*> (object discovery): Questions about an entity (the subject) with specific relation constraints (the predicate) but misses the target entity (the object). The task is to answer one specific aspect of the entity. Example: “What theories and principles has Isaac Newton developed?”
• <s,*,o> (predicate discovery): Questions about any relations (the predicate) between two entities (the subject and object). The task is to provide the relations between two entities. Example: “How is Isaac Newton and Albert Einstein related?”
• <s,p,o> (fact check): Questions about specific relations (the predicate) between two entities (the subject and object). The task is to check the existence of a specific relationship between the two entities. Example: “Have Isaac Newton and Albert Einstein both contributed to the same same field of science?”
• nested types: Complex questions involving the nesting of multiple basic questions. For example, nested questions can be about general information of an unknown entity with specific relation constraints. Example: “Tell me about the scientist who developed univsersal gravitation.”

Dataset Sources

• Repository: https://github.com/Liu-rj/PolyG
• Paper: https://arxiv.org/pdf/2504.02112
• Graph files: https://drive.google.com/drive/folders/1DJIgRZ3G-TOf7h0-Xub5_sE4slBUEqy9 (from GRBench)

We select the physics graph from the academia domain, goodreads graph from the literature domain, and amazon from the e-commerce domain.

Usage

To preprocess the downloaded graph files, see https://github.com/Liu-rj/PolyG.

Use the question set from each domain:

from datasets import load_dataset
domain = "physics" # can be selected from [physics (academia), goodreads (literature), amazon (e-commerce)]
dataset = load_dataset("Liu-rj/PolyBench", domain, split="test")

Dataset Structure

Information about question generation can be found here: https://github.com/Liu-rj/PolyG/tree/main/benchmarks.

Citation

@misc{liu2025polygadaptivegraphtraversal,
      title={PolyG: Adaptive Graph Traversal for Diverse GraphRAG Questions}, 
      author={Renjie Liu and Haitian Jiang and Xiao Yan and Bo Tang and Jinyang Li},
      year={2025},
      eprint={2504.02112},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2504.02112}, 
}

Dataset Card Authors

Renjie Liu

Dataset Card Contact

[email protected]