flembench — Belgian Dutch vs Netherlands Dutch vocabulary

788 multiple-choice questions (394 matched pairs) that measure whether a language model knows Belgian-Dutch words as well as Netherlands-Dutch words of the same difficulty for humans.

Companion repository, code and full results: github.com/GillesVdSpiegel/flemish-llm-benchmark.

What makes this different

Existing Dutch evaluations (EuroEval, DUMB, the Amsterdam government benchmark) do not distinguish the two varieties. The problem with any such comparison is confounding: if the Belgian questions are simply harder, a lower score proves nothing.

Here, every Belgian word is paired with a Netherlands word that is equally well known in its own country, using word prevalence norms measured separately for Belgium and the Netherlands (Brysbaert et al. 2019). The mean difference within a pair is 0.6 percentage points. A difference in model scores is therefore a property of the model, not of the word list.

Example

{
  "id": "a1a-0437-be",
  "pair_id": "a1a-0437",
  "variety": "be",
  "marker": "velo",
  "prompt": "Wat betekent het woord \"velo\" zoals het in België gebruikt wordt?",
  "choices": ["fiets", "treuzelen", "hal, entree", "salarisschaal"],
  "gold": "A",
  "subcategory": "A1a-clean"
}

The Netherlands member of the pair uses the identical template ("… zoals het in Nederland gebruikt wordt?"), and the correct answer occupies the same position in both, so a model's preference for a particular letter cannot produce a gap.

Fields

Field	Meaning
id, pair_id	item id; both members of a pair share pair_id
variety	be (Belgian Dutch) or nl (Netherlands Dutch)
marker	the word being tested
prompt, choices, gold	question, four options, correct letter
subcategory	A1a-clean (602) or A1a-overlap (186, meaning resembles the word)
register	standard, tussentaal or unspecified
drafted_by, reviewed_by, provenance	where the correct answer came from
split	always public here; 20% of pairs are held out privately

How to use

from datasets import load_dataset

ds = load_dataset("Goomey/flembench", split="train")
item = ds[0]
prompt = item["prompt"] + "\n\n" + "\n".join(
    f"{letter}. {choice}" for letter, choice in zip("ABCD", item["choices"])
)
# ask your model for a single letter, compare with item["gold"]

Report accuracy per variety, and the gap as the mean over pairs of (Belgian correct − Netherlands correct). Use subcategory == "A1a-clean" for the headline figure.

Baseline results (September 2026)

Model	Belgian	Netherlands	Gap
Gemini 3.8 Flash	98.7%	100.0%	−1.3
Claude Sonnet 5	96.8%	98.4%	−1.6
GPT-5.6 Terra	92.5%	96.0%	−3.5
EuroLLM-9B	88.2%	89.8%	−1.6
GEITje-7B-ultra	78.8%	85.3%	−6.4
ChocoLlama-8B	69.2%	59.5%	+9.7
Gemma 4 12B	63.3%	63.0%	+0.3

Computed on all 493 pairs (public plus held-out), clean subset, first run. Only ChocoLlama's gap survives Holm correction for seven models. ChocoLlama is the only model with Belgian training data and the only one that scores better on Belgian Dutch.

How it was built

1. Word selection from the Dutch Crowdsourcing Project norms: words known by ≥80% in their own country and ≥30 points less in the other. 584 Belgian, 497 Netherlands candidates.
2. Pairing: one-to-one matching on own-country familiarity (±3 points) and on the between-country difference (±5 points). 493 pairs.
3. Correct answers: drafted independently by Claude Opus 5 and Gemini. 1,018 of 1,081 agreed; the 63 disagreements were decided by the author (a native Flemish speaker); 11 uncertain Netherlands words were decided by a Netherlands-Dutch speaker, who rejected 4.
4. Distractors: meanings of other words in the set, with no shared content words with the correct answer, no resemblance to the target word, and comparable length.

Provenance of the correct answers in this public split: 762 drafted by Claude, 17 by Gemini, 4 written by the author, 5 by the Netherlands reviewer; 194 items carry an explicit human review.

Limitations

• Isolated word meaning only. No context sentences, no syntax, no administrative language, no institutional knowledge, no spoken language.
• Ceiling effect. The strongest models score 97–100%, leaving little room to measure a difference; their gaps are lower bounds.
• Correct answers are partly LLM-drafted, and one of the drafting models (Claude) is among the evaluated models. An advantage for Claude cannot be ruled out. Use drafted_by to check.
• No human baseline. Model scores are not compared against native speakers answering the same questions.
• Noise floor. Repeated runs give 93–99% identical answers, so differences under 1–2 points can be run-to-run variation.
• A1a-overlap items are easier; exclude them for the headline.
• Single author for the Belgian side; no second Flemish reviewer has checked a sample yet.

Licence and citation

CC BY 4.0.

@misc{flembench2026,
  title  = {flembench: a paired benchmark for Belgian vs Netherlands Dutch vocabulary in LLMs},
  author = {Gilles (@GillesVdSpiegel)},
  year   = {2026},
  url    = {https://github.com/GillesVdSpiegel/flemish-llm-benchmark}
}

Word selection derives from the prevalence norms of Brysbaert, Keuleers, Mandera & Stevens (2019), Recognition Times for 54 Thousand Dutch Words, Psychologica Belgica 59(1), licensed CC BY-NC 4.0. Those norms are not redistributed in this dataset; only words, questions and our own meanings are included.