ArbelTepper/extended_test_xnli

XNLI-MT: the XNLI test set extended from 14 to 48 languages

This dataset extends facebook/xnli with 34 additional languages.

The XNLI test set (5,010 premise/hypothesis pairs) for 48 languages (+ English): the 14 original XNLI languages and 34 additional languages, chosen as the FLORES+ languages used in our multilingual benchmarks that XNLI does not cover. Gold labels are the same for every language (inherited from English).

Where each language comes from

• The 14 languages that are also in facebook/xnli (ar bg de el es fr hi ru sw th tr ur vi zh) are NOT translated with TranslateGemma. Their test.jsonl files are the test split of facebook/xnli for that language, copied unchanged (human translations); we only added the English originals and the translation_source column.
• The 34 new languages are machine-translated from the English XNLI test set with google/translategemma-4b-it (English -> target, greedy decoding, official chat template, no human post-editing). The languages added in this dataset are: Bengali, Catalan, Croatian, Czech, Danish, Dutch, Estonian, Filipino, Finnish, Gujarati, Hebrew, Hungarian, Icelandic, Indonesian, Italian, Japanese, Kannada, Korean, Latvian, Lithuanian, Malayalam, Marathi, Norwegian, Persian, Polish, Portuguese, Romanian, Serbian, Slovak, Slovenian, Swedish, Tamil, Telugu, Ukrainian.

Every row carries a translation_source field (facebook/xnli, google/translategemma-4b-it or original (English)).

Structure

One folder/config per language: <lang>/test.jsonl; en/test.jsonl is the English source.

from datasets import load_dataset
ds = load_dataset("ArbelTepper/extended_test_xnli", "bn", split="test")

Fields: idx (row id in the English test set), premise, hypothesis (in the language), premise_en, hypothesis_en, premise_lang, hypothesis_lang, label (0 entailment, 1 neutral, 2 contradiction), full_label, translation_source.

Length metadata (length_metadata.json)

length_metadata.json maps each machine-translated language code (the 34 TranslateGemma languages; the 14 languages copied from facebook/xnli have no entry) to the list of rows (idx values, as in <lang>/test.jsonl) whose translation length looks unusual compared with the English source.

For every translated text the check computes ln(characters in the translation / characters in the English source) and flags it when this is more than 4 standard deviations from that language's mean, which catches dropped content, padding and runaway output. A row is listed if its premise or its hypothesis was flagged (open the row to see which one); each premise is shared by three consecutive rows, so a flagged premise lists all three of them. 261 rows are listed across all languages (0.15% of 170,340).

Other checks on every translated text found no problems in any language: empty output, output identical to the English source, truncated generation (the 512-token limit), repetition loops, meta text such as "Here is the translation", and missing translations. The first run produced 38 truncated or looping texts; they were re-translated with a repetition penalty (1.1, then 1.3 for the 7 that remained), and none are left.

Languages

folder	language	source	FLORES code	TranslateGemma code
ar	Arabic	facebook/xnli	-	-
bg	Bulgarian	facebook/xnli	-	-
bn	Bengali	TranslateGemma-4B	ben_Beng	bn
ca	Catalan	TranslateGemma-4B	cat_Latn	ca
cs	Czech	TranslateGemma-4B	ces_Latn	cs
da	Danish	TranslateGemma-4B	dan_Latn	da
de	German	facebook/xnli	-	-
el	Greek	facebook/xnli	-	-
es	Spanish	facebook/xnli	-	-
et	Estonian	TranslateGemma-4B	est_Latn	et
fa	Persian	TranslateGemma-4B	pes_Arab	fa
fi	Finnish	TranslateGemma-4B	fin_Latn	fi
fil	Filipino	TranslateGemma-4B	fil_Latn	fil-PH
fr	French	facebook/xnli	-	-
gu	Gujarati	TranslateGemma-4B	guj_Gujr	gu
he	Hebrew	TranslateGemma-4B	heb_Hebr	he
hi	Hindi	facebook/xnli	-	-
hr	Croatian	TranslateGemma-4B	hrv_Latn	hr
hu	Hungarian	TranslateGemma-4B	hun_Latn	hu
id	Indonesian	TranslateGemma-4B	ind_Latn	id
is	Icelandic	TranslateGemma-4B	isl_Latn	is
it	Italian	TranslateGemma-4B	ita_Latn	it
ja	Japanese	TranslateGemma-4B	jpn_Jpan	ja
kn	Kannada	TranslateGemma-4B	kan_Knda	kn
ko	Korean	TranslateGemma-4B	kor_Hang	ko
lt	Lithuanian	TranslateGemma-4B	lit_Latn	lt
lv	Latvian	TranslateGemma-4B	lvs_Latn	lv
ml	Malayalam	TranslateGemma-4B	mal_Mlym	ml
mr	Marathi	TranslateGemma-4B	mar_Deva	mr
nl	Dutch	TranslateGemma-4B	nld_Latn	nl
no	Norwegian	TranslateGemma-4B	nob_Latn	no
pl	Polish	TranslateGemma-4B	pol_Latn	pl
pt	Portuguese	TranslateGemma-4B	por_Latn	pt-BR
ro	Romanian	TranslateGemma-4B	ron_Latn	ro
ru	Russian	facebook/xnli	-	-
sk	Slovak	TranslateGemma-4B	slk_Latn	sk
sl	Slovenian	TranslateGemma-4B	slv_Latn	sl
sr	Serbian	TranslateGemma-4B	srp_Cyrl	sr-Cyrl
sv	Swedish	TranslateGemma-4B	swe_Latn	sv
sw	Swahili	facebook/xnli	-	-
ta	Tamil	TranslateGemma-4B	tam_Taml	ta
te	Telugu	TranslateGemma-4B	tel_Telu	te
th	Thai	facebook/xnli	-	-
tr	Turkish	facebook/xnli	-	-
uk	Ukrainian	TranslateGemma-4B	ukr_Cyrl	uk
ur	Urdu	facebook/xnli	-	-
vi	Vietnamese	facebook/xnli	-	-
zh	Chinese	facebook/xnli	-	-

Limitations

• The 34 new languages are fully machine-translated and not human-verified.
• Premise and hypothesis are translated independently.
• Rows whose translation length looks unusual are listed in length_metadata.json (see above); they were not corrected.

License and terms

The 14 original languages and the English source come from XNLI/MultiNLI (verify the upstream license; XNLI is CC BY-NC 4.0). The machine-translated languages are derived from them and generated with a Gemma model, so the Gemma Terms of Use apply to those files. Please check both before reuse.

Citation

If you use this dataset, please cite it together with XNLI and TranslateGemma:

@misc{tepper2026extendedxnli,
  title        = {{XNLI} Test Set Extended to 48 Languages},
  author       = {Tepper, Arbel},
  year         = {2026},
  howpublished = {Hugging Face dataset, \url{https://huggingface.co/datasets/ArbelTepper/extended_test_xnli}}
}

@inproceedings{conneau2018xnli,
  title     = {{XNLI}: Evaluating Cross-lingual Sentence Representations},
  author    = {Conneau, Alexis and Rinott, Ruty and Lample, Guillaume and Williams, Adina and Bowman, Samuel R. and Schwenk, Holger and Stoyanov, Veselin},
  booktitle = {Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing},
  year      = {2018},
  publisher = {Association for Computational Linguistics}
}

@misc{finkelstein2026translategemma,
  title         = {{TranslateGemma} Technical Report},
  author        = {Finkelstein, Mara and Caswell, Isaac and Domhan, Tobias and others},
  year          = {2026},
  eprint        = {2601.09012},
  archivePrefix = {arXiv}
}