You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Spectrum multilingual SFT

Multilingual supervised fine-tuning data for Spectrum (a reproduction of OmniSONAR, Sec. 9). Every example has its context in language X and its question, choices and answer in English, which is the X->En format of Spectrum's evaluation. The English-only mix is in VardanSharma13/spectrum-sft; this set does not repeat it.

config what languages train validation test
silver English SFT contexts machine-translated with NLLB-200-3.3B 201 1,769,291 54,331 46,330
native real QA written in the language (native passages); Q/A translated to English 25 317,857 19,174 19,165

silver: translated contexts, 201 languages

  • Source: the released English SFT mix (spectrum-sft), contexts of at most 12 sentences and 300 words, at most 600 (phase 1) or 150 (phase 2) examples per dataset per language, a different random subset per language.
  • 56 languages x ~21k examples (phase 1) and the other 145 NLLB-200 languages x 5.4k (phase 2).
  • Translation: facebook/nllb-200-3.3B, sentence by sentence (NLTK punkt), beam 2; paragraphs kept.
  • Quality filter: an example is dropped if any sentence is empty, cut at the length limit, more than 3x the source length, a repetition loop, or an untranslated copy, or if GlotLID says the context is English. Kept 1,877,912 of 1,959,000 (95.9%).
  • Columns: the English SFT schema (context (translated), task_type, prompt_question, prompt_choices, target, dataset, split, uid) plus lang (NLLB code), context_en (original), lid (GlotLID label of the translation), source = silver_nllb, format = x_context_en_qa. uid is the English example's uid.
  • For extractive QA the English answer may not appear word for word in the translated context.

Languages that lost the most examples (low-resource NLLB directions; lid_mismatch is the share of kept examples whose GlotLID label differs from the target language: reported, not filtered, since language ID is unreliable for close varieties and scripts GlotLID rarely sees):

language kept lid_mismatch
knc_Arab 932 / 5,400 1.00
quy_Latn 3,486 / 5,400 0.76
dik_Latn 3,510 / 5,400 0.01
aeb_Arab 3,527 / 5,400 0.60
ace_Arab 3,628 / 5,400 0.22
kbp_Latn 3,665 / 5,400 0.01
acm_Arab 3,706 / 5,400 0.97
fon_Latn 3,717 / 5,400 0.03
bjn_Arab 3,828 / 5,400 0.12
knc_Latn 3,870 / 5,400 0.01

native: real multilingual QA, 25 languages

Passages and questions written by native speakers (Wikipedia and news), not translated. For training in the X->En format, the question and answer are translated to English with the same NLLB-200-3.3B (the passage is not); the originals are kept in question_native and answer_native.

  • format = native: native passage, English question, English answer.
  • format = xorqa (XOR-TyDi): native question, English passage, English answer (cross-lingual QA). The translated question is in prompt_question, the original in question_native.
  • answer_type (rule-based from the English question and answer): name 103,853, number 49,240, date 46,726, other 156,377.
  • Checks: numbers must survive translation (the digits of the native answer, in any script, must equal the digits of the English answer; purely numeric answers are copied in ASCII digits); translations cut at the length limit, untranslated copies and looping answers are dropped.
  • Splits: source train -> train; source validation -> validation / test (half each, by passage). Test splits of the source datasets are never used. Train passages that also occur in validation/test are dropped.
  • Decontamination: passages sharing a 13-word (or 25-character, for Chinese/Japanese/Thai) sequence with any evaluation set are removed from all splits: Belebele, SIB-200, MLQA, XQuAD, IndicQA, IndicGenBench XQuAD-IN and XorQA-IN, ECLeKTic, MultiLoKo.
dropped rows
duplicate_context_question 1,193
digits_changed_in_translation 17,142
answer_translation_too_long 6,206
bad_answer_translation 669
question_translation_too_long 36
bad_question_translation 41
train_context_in_val_test 28,343
eval_contamination 6,218
language train validation test
fas 51,703 3,206 3,131
pol 43,689 2,709 2,759
rus 34,488 3,070 2,910
jpn 28,592 1,757 1,876
ces 24,713 0 0
vie 17,492 1,089 1,263
arb 16,225 855 923
spa 14,620 918 897
cat 11,093 749 721
deu 10,922 0 0
swh 10,071 246 238
fin 8,262 581 529
cmn 5,992 1,287 1,170
heb 7,145 230 189
tel 5,994 338 349
ind 5,475 270 279
tur 3,173 1,035 1,037
tha 4,329 240 323
kor 3,497 277 267
ben 3,633 74 79
fra 3,722 0 0
nob 2,056 243 225
est 487 0 0
hin 305 0 0
tam 179 0 0

Sources (licences as stated on each dataset's card; please follow them and cite the original work):

source dataset languages rows licence
pquad Gholamreza/pquad fas 51,410 cc-by-sa-4.0
poquad clarin-pl/poquad pol 49,157 cc-by-4.0
tydiqa google-research-datasets/tydiqa arb, ben, fin, ind, kor, rus, swh, tel 46,555 apache-2.0
sberquad kuznetsoffandrey/sberquad rus 31,518 unknown
jaquad SkelterLabsInc/JaQuAD jpn 30,082 cc-by-sa-3.0
sqad_cs CZLC/SQAD_3.2 ces 24,713 not stated
viquad taidng/UIT-ViQuAD2.0 vie 19,844 not stated
sqac PlanTL-GOB-ES/SQAC spa 16,435 cc-by-sa-4.0
xor_tydi_rc coastalcph/tydi_xor_rc arb, ben, fin, jpn, kor, rus, tel 14,293 mit
viquiquad projecte-aina/viquiquad cat 12,563 cc-by-sa-4.0
germanquad deepset/germanquad deu 10,922 cc-by-4.0
cmrc2018 hfl/cmrc2018 cmn 8,449 cc-by-sa-4.0
heq Etelis/HeQ_v1 heb 7,564 cc-by-4.0
kenswquad Kencorpus/KenSwQuAD swh 7,492 cc-by-4.0
persianqa SajjadAyoubi/persian_qa fas 6,630 not stated
tquad2 boun-tabilab/TQuad-2 tur 5,245 not stated
iapp iapp/iapp_wiki_qa_squad tha 4,892 mit
piaf AgentPublic/piaf fra 3,722 mit
norquad ltg/norquad nob 2,524 cc0-1.0
arcd hsseinmz/arcd arb 1,215 mit
estqa TalTechNLP/EstQA est 487 cc-by-4.0
chaii nirantk/chaii-hindi-and-tamil-question-answering hin, tam 484 not stated

Not used: KorQuAD and IndoQA (no-derivatives licences), ThaiQA (non-commercial), and every evaluation set above.

Licence

license: other. Each row keeps the licence of its source dataset (English sources: see spectrum-sft; native sources: table above). The translations were produced with NLLB-200-3.3B, whose weights are released under CC BY-NC 4.0, so use this data for non-commercial research.

Downloads last month
233