EuroMultiSynth-Instruct: synthetic instruction-tuning data in 35 European languages

4,160,598 single-turn instruction/response pairs in 35 European (and neighbouring) languages, written natively in each language (not translated) by an LLM, scored by an LLM judge, quality-filtered and de-duplicated within each language. Built with synthgen for the OpenEuroLLM project.

Quick start

from datasets import load_dataset

ds = load_dataset("openeurollm/EuroMultiSynth-Instruct", "de", split="train")   # one language
ds = load_dataset("openeurollm/EuroMultiSynth-Instruct", split="train")         # all 35 languages
print(ds[0]["messages"])

Fields

field	type	description
id	string	unique row id, <lang>-<number>
language	string	ISO 639-1 code
messages	list of {role, content}	[user prompt, assistant response]
domain	string	task category the prompt was generated for (18 values)
persona	string or null	user persona the prompt was written as (null for 87% of rows)
judge_score	int8	LLM-judge quality score, 0-10 (only 7-10 kept)

How it was made

1. Generation. google/gemma-4-31b-it (served with vLLM) wrote each instruction and its response in the target language in one call, conditioned on a task domain and, for some prompts, a user persona. 7,040,000 unique samples were generated (160k or 320k per language).
2. Judging. Qwen/Qwen3.6-27B scored the samples 0-10 (temperature 0, thinking disabled). 2,204 samples (0.03%) were never judged and are not included.
3. Quality filter. Kept samples with score ≥ 7: 5,688,783 rows (80.9% of scored samples).
4. Exact de-duplication on (language, whitespace/case-normalised prompt): −93,387 rows.
5. Near-duplicate removal, within each language only. Prompts were embedded and a prompt was dropped if its cosine similarity to an earlier kept prompt in the same language was at or above the threshold (greedy, first occurrence kept). 31 languages used sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 at 0.97; Maltese, Irish, Basque and Icelandic, which that model does not cover, used sentence-transformers/LaBSE at 0.96 (calibrated on German and Croatian to remove the same share as MiniLM at 0.97). Thresholds were chosen from sweeps with example pairs per similarity band: at these values, removed pairs are rewordings of the same request. −1,434,798 rows.

Final: 4,160,598 rows.

Languages

code	language	rows	near-dup model / threshold
tr	Turkish	236,352	MiniLM ≥ 0.97
de	German	230,459	MiniLM ≥ 0.97
bg	Bulgarian	212,388	MiniLM ≥ 0.97
bs	Bosnian	207,040	MiniLM ≥ 0.97
gl	Galician	193,629	MiniLM ≥ 0.97
hr	Croatian	190,139	MiniLM ≥ 0.97
da	Danish	185,955	MiniLM ≥ 0.97
ca	Catalan	166,246	MiniLM ≥ 0.97
fi	Finnish	132,641	MiniLM ≥ 0.97
pl	Polish	126,767	MiniLM ≥ 0.97
it	Italian	124,420	MiniLM ≥ 0.97
ro	Romanian	119,021	MiniLM ≥ 0.97
no	Norwegian	115,400	MiniLM ≥ 0.97
mk	Macedonian	110,730	MiniLM ≥ 0.97
sr	Serbian	110,270	MiniLM ≥ 0.97
uk	Ukrainian	109,240	MiniLM ≥ 0.97
sq	Albanian	107,505	MiniLM ≥ 0.97
sk	Slovak	106,467	MiniLM ≥ 0.97
sv	Swedish	104,862	MiniLM ≥ 0.97
pt	Portuguese	104,727	MiniLM ≥ 0.97
fr	French	103,013	MiniLM ≥ 0.97
is	Icelandic	102,362	LaBSE ≥ 0.96
nl	Dutch	102,213	MiniLM ≥ 0.97
cs	Czech	100,099	MiniLM ≥ 0.97
es	Spanish	99,709	MiniLM ≥ 0.97
el	Greek	99,005	MiniLM ≥ 0.97
ka	Georgian	90,665	MiniLM ≥ 0.97
sl	Slovenian	85,119	MiniLM ≥ 0.97
lt	Lithuanian	71,445	MiniLM ≥ 0.97
hu	Hungarian	68,390	MiniLM ≥ 0.97
lv	Latvian	62,438	MiniLM ≥ 0.97
mt	Maltese	52,674	LaBSE ≥ 0.96
et	Estonian	52,551	MiniLM ≥ 0.97
eu	Basque	50,716	LaBSE ≥ 0.96
ga	Irish	25,941	LaBSE ≥ 0.96

Domains

domain	rows	share
general question answering	855,672	20.6%
coding	585,753	14.1%
brainstorming & idea generation	555,206	13.3%
reports, summaries & documentation	287,952	6.9%
roleplay & character play	223,106	5.4%
translation	195,296	4.7%
math (word problems & explanations)	179,613	4.3%
short stories & fiction	165,221	4.0%
social media & blog posts	136,750	3.3%
data analysis (tables, lists, datasets)	133,886	3.2%
summarization	124,134	3.0%
meetings, planning & project coordination	108,337	2.6%
workplace email & communication	104,977	2.5%
humor, jokes & satire	104,715	2.5%
poetry & song lyrics	104,456	2.5%
customer & client communication	99,965	2.4%
job applications & CVs	99,951	2.4%
personal essays & reflections	95,608	2.3%

Judge scores

score	rows	share
7	283,975	6.8%
8	168,940	4.1%
9	476,087	11.4%
10	3,231,596	77.7%

Personas (8): office worker (69,697), teenager (69,427), retiree (69,304), small-business owner (68,846), recent immigrant (68,527), parent (68,376), university student (67,856), tourist (63,608).

Limitations

• Fully synthetic. Both the prompts and the responses are model-generated; facts, figures and local details (laws, prices, places) can be wrong or out of date.
• LLM-judged. Quality filtering relies on a single LLM judge, whose reliability likely varies by language and is probably weaker for lower-resource languages.
• Uneven size and quality across languages. Languages with lower judge pass rates are smaller (e.g. Irish: 320,000 generated, 25,941 kept). A few spot-checked Basque and Irish samples contained malformed words, so expect more noise in the lower-resource languages.
• Residual redundancy. The near-duplicate thresholds are deliberately conservative, so many prompts still share a template or scenario with different specifics.
• Single-turn only.

License

Released under the Apache License 2.0.

This choice follows from the models used to build the dataset, all of which are Apache-2.0:

role	model	license
generation (wrote every prompt and response)	google/gemma-4-31B-it	Apache-2.0
quality judge (scores only, no text in the dataset)	Qwen/Qwen3.6-27B	Apache-2.0
near-duplicate filtering	sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2, sentence-transformers/LaBSE	Apache-2.0

All text in this dataset is model output, attributed to google/gemma-4-31B-it. Please credit both the dataset and the generating model when you use it, and follow Google's Gemma Prohibited Use Policy in spirit when building on Gemma-generated data.

Citation

TODO