RefusalBench-NQ

RefusalBench-NQ is a single-document diagnostic benchmark for selective refusal in retrieval-augmented generation (RAG): the ability of a grounded language model to answer when the context supports a reliable answer and to refuse with the correct reason when it does not. It is one of the two benchmarks released with the paper RefusalBench: Generative Evaluation of Selective Refusal in Grounded Language Models.

Each instance is produced by applying exactly one of 176 linguistic perturbation levers (across 6 uncertainty classes × 3 intensities) to a verified-answerable Natural Questions item, then filtered through a 4-model generator–verifier pipeline requiring unanimous agreement.

• Paper: EACL 2026 (ACL Anthology) · arXiv preprint
• Code / generation framework: https://github.com/aashiqmuhamed/refusalbench
• Companion benchmark: RefusalBench-GaRAGe (multi-document)
• Size: 1,600 instances (single test split) · 100 source NQ questions

The task

The model is shown perturbed_query + perturbed_context and must either answer, or emit a single refusal code. The ground-truth label is expected_rag_behavior:

• LOW intensity → ANSWER_CORRECTLY (the perturbation is sub-threshold; a competent model should still answer)
• MEDIUM / HIGH intensity → a REFUSE_* code (the defect warrants refusal)

Refusal codes: REFUSE_AMBIGUOUS_QUERY, REFUSE_CONTRADICTORY_CONTEXT, REFUSE_INFO_MISSING_IN_CONTEXT, REFUSE_FALSE_PREMISE_IN_QUERY, REFUSE_GRANULARITY_MISMATCH, REFUSE_NONFACTUAL_QUERY.

Perturbation classes: P-Ambiguity, P-Contradiction, P-MissingInfo, P-FalsePremise, P-GranularityMismatch, P-EpistemicMismatch.

Data fields

field	type	description
id	string	Unique instance id (RB-NQ_…).
source_id	string	Source Natural Questions question id.
generator_model	string	Model that authored the perturbation (claude/deepseek/gpt/nova).
perturbation_class	string	One of the 6 uncertainty classes.
intensity	string	LOW / MEDIUM / HIGH.
expected_rag_behavior	string	Ground-truth label (ANSWER_CORRECTLY or a REFUSE_* code).
lever_selected	string	The specific linguistic lever applied.
original_query	string	Unperturbed query.
original_context	string	Unperturbed (answerable) passage.
original_answers	list[string]	Reference short answers for the original item.
perturbed_query	string	Query shown to the model under test.
perturbed_context	string	Context shown to the model under test.
implementation_reasoning	string	Generator's rationale for how the perturbation was built.
verifier_votes	dict	Per-verifier PASS/FAIL (all rows are unanimous PASS).

implementation_reasoning is an answer key. It explains how/why the perturbation was introduced and often reveals the expected behavior. Do not include it in the prompt to the model under test — feed only perturbed_query and perturbed_context. It is provided for transparency and analysis.

Usage

from datasets import load_dataset

ds = load_dataset("aashiqmuhamed/RefusalBench-NQ", split="test")
ex = ds[0]
# Model input: ex["perturbed_query"], ex["perturbed_context"]
# Ground truth: ex["expected_rag_behavior"]

Construction & quality control

A base set of 100 NQ questions that all evaluated frontier models answered correctly in their unperturbed state was perturbed with the 176-lever engine. Every candidate was assessed by four verifier models (Claude-4-Sonnet, DeepSeek-R1, GPT-4o, Nova-Pro); only unanimously approved perturbations are included. Stratified sampling balances the final set across perturbation types and the four generators (exactly 400 per generator). Expert human validation of a 180-item stratified sample gave a 93.1% pass rate.

Properties constant across all rows (recorded here rather than per row): filtering = unanimous consensus; verifiers = 4; sampling = stratified.

Licensing & attribution

Released under Apache-2.0, inherited from the source datasets: Natural Questions (Kwiatkowski et al., 2019) and KILT (Petroni et al., 2020). Please cite those works in addition to RefusalBench. The RefusalBench code/framework is also Apache-2.0; see the GitHub repository.

Citation

@inproceedings{muhamed-etal-2026-refusalbench,
    title = "{R}efusal{B}ench: Generative Evaluation of Selective Refusal in Grounded Language Models",
    author = "Muhamed, Aashiq  and
      Ribeiro, Leonardo F. R.  and
      Dreyer, Markus  and
      Smith, Virginia  and
      Diab, Mona T.",
    booktitle = "Proceedings of the 19th Conference of the {E}uropean Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.eacl-long.321/",
    doi = "10.18653/v1/2026.eacl-long.321",
    pages = "6811--6856"
}