Dataset Card for HalluMat

Paper | Code

Dataset Summary

HalluMat is a benchmark for measuring hallucination in LLM answers to materials science questions. It has 655 materials science questions, each asked in several paraphrased forms, for 3,269 question and answer pairs in total. Every answer was generated by Llama-2-7B-chat and compared with a reference answer, then labelled Low, Medium or High hallucination.

Because each question is asked several ways, you can measure not only whether a model is wrong but whether it stays consistent when the same question is rephrased. The paper calls this the Paraphrased Hallucination Consistency Score (PHCS).

Supported Tasks

• Hallucination detection: predict hallucination_level from the question, the generated response and the reference answer.
• Consistency analysis: use group_id to compare answers to paraphrases of the same question.

Languages

English.

Dataset Structure

Data Instances

{
  "id": 5,
  "group_id": 2,
  "question": "In which group do steels, specifically iron and carbon, belong?",
  "generated_response": "In which group do steels, specifically iron and carbon, belong?\n\nSteels are a type of alloy that contains iron and carbon, ...",
  "reference_answer": "Metallic Solid",
  "hallucination_score": 0.3018,
  "hallucination_level": "High",
  "similarity_score": 0.0,
  "recomputed_level": "High"
}

Data Fields

responses config (one row per generated answer):

Field	Type	Description
id	int	Row id
group_id	int	The original question. Rows with the same group_id are paraphrases of one question
question	string	The paraphrased question given to the model
generated_response	string	Raw Llama-2-7B-chat output. It usually starts by repeating the question
reference_answer	string	The verified answer (missing for 4 rows)
hallucination_score	float	Hallucination score from the HalluMat pipeline (higher means more hallucination)
hallucination_level	string	Low, Medium or High, the label used in the paper
similarity_score	float	Embedding similarity between the response and the reference answer
recomputed_level	string	The level recomputed from similarity_score alone, kept for comparison

groups config (one row per question):

Field	Type	Description
group_id	int	The original question
n_low, n_medium, n_high	int	Number of paraphrases in each hallucination level
score_mean, score_std, score_min, score_max	float	Summary of hallucination_score across the paraphrases

Data Splits

There is a single test split.

	Count
Questions (groups)	655
Question and answer pairs	3,269
Low hallucination	57
Medium hallucination	872
High hallucination	2,340

from datasets import load_dataset

responses = load_dataset("bhanuprakashvangala/HalluMat", "responses", split="test")
groups = load_dataset("bhanuprakashvangala/HalluMat", "groups", split="test")

Dataset Creation

Curation Rationale

General hallucination benchmarks rarely cover scientific domains, where a wrong answer can mislead research. HalluMat focuses on materials science and adds paraphrases so that consistency can be measured along with correctness.

Source Data

Materials science questions with verified answers were paraphrased into several equivalent forms. Llama-2-7B-chat answered each paraphrase, and each answer was scored against the reference answer with the HalluMat pipeline. The code is in the HalluMat repository.

Changes from the Original Release

The original CSV had 6 more rows with an empty question and response. They were counted as High in the paper (2,346 High) and are removed here, which gives the 3,269 pairs reported in the paper and 2,340 High. Column names were changed to lowercase snake case. Values are otherwise unchanged.

Personal and Sensitive Information

The data contains only scientific questions and model outputs. It has no personal information.

Limitations

• Labels are produced automatically, not by human annotators.
• generated_response is raw model output and often repeats the question before answering.
• The labels are skewed toward High, so report per-class or macro metrics.

Licensing Information

Released under CC BY 4.0.

Citation

@misc{vangala2025hallumatdetectinghallucinationsllmgenerated,
      title={HalluMat: Detecting Hallucinations in LLM-Generated Materials Science Content Through Multi-Stage Verification}, 
      author={Bhanu Prakash Vangala and Sajid Mahmud and Pawan Neupane and Joel Selvaraj and Jianlin Cheng},
      year={2025},
      eprint={2512.22396},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2512.22396}, 
}

Contact

Please open a discussion on this dataset page or an issue in the code repository.