themohal/pakistan-languages-dataset

Pakistan Languages Dataset

A parallel dataset of English sentences translated into ~37 languages of Pakistan. One row per English sentence; one column per language, using its ISO 639-3 code.

⚠️ Translation quality varies a lot by language — read this before using the data

Every language column is force-filled: no cell is ever left blank. For high-resource languages this just means a verified translation. For the lowest-resource languages, it can mean best guess at a language it has very little or no training data for — possibly leaning on a related language, a partial/word-by-word construction, or transliteration rather than a fluent, checked translation. Mistakes, awkward phrasing, and outright errors should be expected in those columns.

Two per-row columns tell you exactly which cells to trust less:

• low_confidence_languages — comma-separated ISO codes of languages in that row that either never cleared the quality judge (score ≥ 90, or ≥ 80 plus a passing repair pass) or were never attempted by the bulk translation call and had to be force-generated afterward.
• low_confidence_count — how many of the 37 language columns in that row fall into the above category.
• coverage_count / filled_languages — the complement: how many, and which, languages in that row are high-confidence (cleared the judge threshold).

Always filter on low_confidence_languages per row rather than trusting the static tier below. Tier is a general expectation, not a per-row guarantee — an otherwise "high" tier language can occasionally land in a row's low_confidence_languages if that particular sentence was hard (idioms, technical jargon, etc.), and a "low" or "experimental" language can occasionally score well.

Language coverage tiers

These tiers describe how reliable each language is expected to be on average, based on how much training data realistically has for it. They are informational only — every language is attempted for every row regardless of tier.

High — genuinely fluent (verify occasional idioms/jargon only)

Code	Language	Script
urd	Urdu	Perso-Arabic (Nastaliq)
pnb	Punjabi	Perso-Arabic (Shahmukhi)
pbu	Pashto	Perso-Arabic
snd	Sindhi	Perso-Arabic

Mid — workable quality, spot-check before trusting fully

Code	Language	Script
skr	Saraiki	Perso-Arabic (Shahmukhi)
bal	Balochi	Perso-Arabic
hno	Hindko	Perso-Arabic
phr	Pahari-Pothwari	Perso-Arabic
brh	Brahui	Perso-Arabic
kas	Kashmiri	Perso-Arabic

Low — sample-able but inconsistent; spot-check before any real use

Code	Language	Script
gju	Gojri	Perso-Arabic
scl	Shina	Perso-Arabic
khw	Khowar	Perso-Arabic
bft	Balti	Perso-Arabic / Tibetan
bsk	Burushaski	Perso-Arabic
wbl	Wakhi	Perso-Arabic
trw	Torwali	Perso-Arabic
kls	Kalasha	Latin / Perso-Arabic
gwc	Gawri (Kalami)	Perso-Arabic
phl	Palula	Perso-Arabic

Experimental — expect frequent errors; treat as best-effort guesses, not ground truth

Little to no digital text exists for most of these languages anywhere. Translations here are the weakest link in the dataset and are the most likely to be inaccurate, partially transliterated, or borrowed from a related language.

Code	Language	Script
dml	Dameli	Perso-Arabic
ush	Ushojo	Perso-Arabic
ydg	Yidgha	Perso-Arabic
dmk	Domaaki	Perso-Arabic
btv	Bateri	Perso-Arabic
clh	Chilisso	Perso-Arabic
gwf	Gowro	Perso-Arabic
xka	Kalkoti	Perso-Arabic
shd	Kundal Shahi	Perso-Arabic
oru	Ormuri	Perso-Arabic
wne	Wanetsi (Waneci)	Perso-Arabic
mki	Dhatki	Perso-Arabic / Devanagari
jdg	Jadgali	Perso-Arabic
xhe	Khetrani	Perso-Arabic
bdz	Badeshi	Perso-Arabic
aeq	Aer	Perso-Arabic
haz	Hazaragi	Perso-Arabic

ISO codes are best-effort — several of these languages have no single agreed-upon code. Verify against Ethnologue/Glottolog before treating a code as authoritative for publication.

Recommended usage

• Training/fine-tuning: filter out rows/cells listed in low_confidence_languages, or restrict to the High and Mid tiers, unless you specifically want (and can tolerate noise in) Low/Experimental coverage.
• Evaluation: do not use Low/Experimental-tier cells as gold references without native-speaker review.
• Research on low-resource MT: the Low/Experimental columns may still be useful as a starting point for further human correction, but should not be presented as verified translations as-is.

Columns

Column	Description
english	Source English sentence
tier	Length/complexity tier of the English sentence (not language confidence)
domain	Topic/register of the English sentence
coverage_count	Number of languages that cleared the quality judge for this row
filled_languages	ISO codes of the high-confidence languages for this row
low_confidence_count	Number of languages in this row that are best-effort/unverified
low_confidence_languages	ISO codes of the best-effort/unverified languages for this row
<iso_code> (×37)	Translation text for that language, always non-empty