SommBench

SommBench is a multilingual benchmark for assessing sommelier expertise in large language models. It comprises 3,024 expert-curated examples across eight languages (da, de, en, es, fi, it, sk, sv), designed by professional sommeliers to evaluate sensory grounding, factual wine knowledge, and practical pairing skills through three tasks: WTQA, WFC, and FWP.

Configs

wtqa — Wine Theory Question Answering (1,024 examples)

Multiple-choice questions about wine knowledge at varying difficulty levels.

Column	Type	Description
question	string	The question text
a	string	Answer choice A
b	string	Answer choice B
c	string	Answer choice C
d	string	Answer choice D
level	string	Difficulty level (Level 1–4)
true_label	string	Correct answer key (a/b/c/d)
language	string	Language code (da, de, en, es, fi, it, sk, sv)

wfc — Wine Feature Completion (1,000 examples)

Complete missing attributes of a wine profile from partial metadata using structured generation.

Column	Type	Description
url	string	Source URL
title	string	Wine name/vintage
type	string	Wine type (red, white, …)
sugar	float64	Sugar content (g/L)
alcohol	float64	Alcohol percentage
country	string	Country of origin
region	list[string]	Wine region(s)
grapes	list[string]	Grape variety/varieties
dryness	string	Dryness classification
acidity	string	Acidity classification
body	string	Body classification

fwp — Food-Wine Pairing (1,000 examples)

Binary classification of whether a wine pairs well with a given recipe.

Column	Type	Description
wine	string	Wine name
recipe	string	Recipe / dish description
true_label	string	Pairing label (yes/no)
wine_url	string	Vivino/Alko URL for the wine

Usage

from datasets import load_dataset

wtqa = load_dataset("sommify/test", "wtqa", split="test")
wfc  = load_dataset("sommify/test", "wfc",  split="test")
fwp  = load_dataset("sommify/test", "fwp",  split="test")

Citation

If you use SommBench in your research, please cite:

@misc{brach2026sommbenchassessingsommelierexpertise,
      title={SommBench: Assessing Sommelier Expertise of Language Models},
      author={William Brach and Tomas Bedej and Jacob Nielsen and Jacob Pichna and Juraj Bedej and Eemeli Saarensilta and Julie Dupouy and Gianluca Barmina and Andrea Blasi Núñez and Peter Schneider-Kamp and Kristian Košťál and Michal Ries and Lukas Galke Poech},
      year={2026},
      eprint={2603.12117},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2603.12117}
}

Links

• Paper: SommBench: Assessing Sommelier Expertise of Language Models
• GitHub: https://github.com/sommify/sommbench