MINTQA: A Multi-Hop Question Answering Benchmark for Evaluating LLMs on New and Tail Knowledge

MINTQA evaluates how well LLMs answer complex multi-hop (1–4 hops) questions that involve unpopular and new knowledge. It has two subsets:

Config	Examples	Dimension	Per-hop type labels
pop (MINTQA-POP)	17,887	Knowledge popularity: unpopular vs. popular facts	R, F
ti (MINTQA-TI)	10,479	Time: new facts (only in the 2024 Wikidata dump) vs. old facts (in both 2021 and 2024)	N (new), O (old)

The underlying knowledge graph (a subset of Wikidata) is available at Sp1der/MintQA-KG.

Usage

from datasets import load_dataset

pop = load_dataset("probejie/MINTQA", "pop", split="test")
ti  = load_dataset("probejie/MINTQA", "ti",  split="test")

print(ti[0])

Fields

Field	Type	Description
question	string	The (multi-hop) question
answer	string	Final answer
type	list[string]	Knowledge label for each hop (R/F in pop, N/O in ti)
num_hops	int	Number of hops (= len(triple))
subquestion	list[string] or null	Decomposed single-hop sub-questions (null for 1-hop questions)
subanswer	list[string] or null	Answers to each sub-question (null for 1-hop questions)
triple	list[list[string]]	Wikidata (subject, relation, object) triples forming the reasoning chain
facts	list[string]	Related knowledge facts about the entities involved

Example (ti):

{
  "question": "Where was the author of \"Structural studies on bacteriophage portal proteins\" educated?",
  "answer": "University of York",
  "type": ["N", "N"],
  "num_hops": 2,
  "subquestion": ["Who is the author of the work titled \"Structural studies on bacteriophage portal proteins\"?", "Where was Weisha Luan educated?"],
  "subanswer": ["Weisha Luan", "University of York"],
  "triple": [["Structural studies on bacteriophage portal proteins", "author", "Weisha Luan"], ["Weisha Luan", "educated at", "University of York"]]
}

The original JSONL files are also available in the GitHub repo.

Citation

@article{he2024mintqa,
  title   = {MINTQA: A Multi-Hop Question Answering Benchmark for Evaluating LLMs on New and Tail Knowledge},
  author  = {He, Jie and Hu, Nan and Long, Wanqiu and Chen, Jiaoyan and Pan, Jeff Z.},
  journal = {arXiv preprint arXiv:2412.17032},
  year    = {2024}
}