Skip to content

SALT-31: A Machine Translation Benchmark Dataset for 31 Ugandan Languages

Sunbird/salt-31

Description

A context-aware MT evaluation benchmark covering 31 Ugandan (and closely related regional) languages. Unlike sentence-level benchmarks (e.g. FLORES), SALT-31 uses five-sentence mini-dialogues built around realistic Ugandan communication scenarios, enabling evaluation of discourse-level phenomena — coreference, register consistency, cultural grounding — in addition to standard translation quality. Used to benchmark Sunbird's Sunflower model and other MT systems.

Languages Covered

31 target languages across three families, plus English as the source language.

Family # Languages
Bantu 17
Nilotic 9
Central Sudanic 5
Code Language Family
eng English Source language
cgg Rukiga Bantu
gwr Lugwere Bantu
kin Kinyarwanda Bantu
koo Rukonjo Bantu
lsm Samia Bantu
lug Luganda Bantu
myx Lumasaba Bantu
nuj Lunyole Bantu
xog Lusoga Bantu
nyn Runyankole Bantu
nyo Runyoro Bantu
rub Lugungu Bantu
ruc Ruruuli Bantu
rwm Kwamba Bantu
swa Swahili Bantu
tlj Lubwisi Bantu
ttj Rutooro Bantu
ach Acholi Nilotic
adh Dhopadhola Nilotic
alz Alur Nilotic
kdi Kumam Nilotic
kdj Karamojong Nilotic
kpz Kupsabiny Nilotic
laj Lango Nilotic
pok Pokot Nilotic
teo Ateso Nilotic
bfa Bari Central Sudanic
keo Kakwa Central Sudanic
lgg Lugbara Central Sudanic
luc Aringa Central Sudanic
mhi Ma'di Central Sudanic

Size / Examples

  • 20 communication scenarios × 5 sentences = 100 English source sentences
  • Each sentence translated into all 31 target languages
  • File Size- 262KB
  • Domains: health, market, family, school, transport, government, agriculture, and other everyday/formal Ugandan contexts

Data Collection Methodology

  • Seed text: Multiple LLMs (GPT-4.5, GPT-4o, DeepSeek R1, LLaMA 3.3 70B, Mistral Large, Gemini 2 Flash, Claude Sonnet) generated candidate English mini-sequences for 20 predefined scenarios; best candidates manually selected for cultural/contextual fit.
  • Translation: Native speakers translated all sentences into the 31 target languages, prioritizing natural phrasing and cultural appropriateness over literal translation.
  • Verification: Independently reviewed by linguists at Makerere University's Department of Linguistics for semantic fidelity, grammar, and cultural appropriateness; disagreements resolved by consensus.

Intended Use Cases

  • Benchmarking MT systems (both directions) across 31 Ugandan languages
  • Diagnostic evaluation of discourse-level translation quality (coreference, register, cultural grounding) — not just sentence-level accuracy
  • Comparing regionally specialized models vs. general-purpose multilingual models

Known Limitation

  • Small (100 sentences) — built for diagnostic evaluation, not model training or fine-grained benchmarking
  • Doesn't cover all sociolinguistic registers, dialects, or code-switching across the 31 languages
  • Text-only for now; speech modality planned for future work

Last Updated

2026-02-02

Citation

Nsumba, S., Akera, B., Ouma, E.N., Ssentanda, M., Kawalya, D., Bainomugisha, E., Mwebaze, E.T., & Quinn, J. (2026). SALT-31: A Machine Translation Benchmark Dataset for 31 Ugandan Languages. Proceedings of the 7th Workshop on African Natural Language Processing (AfricaNLP 2026), pages 211–216. ACL.

hf.co/datasets/Sunbird/salt-31