Skip to content

Evaluating zero-shot translation performance of LLMs for Ghanaian languages

Notifications You must be signed in to change notification settings

GhanaNLP/nsanku

Repository files navigation

Nsanku

A project by Ghana NLP to test the machine translation performance of Large Language Models (LLMs) on Ghanaian languages.

Project Overview

Nsanku is an initiative that evaluates how well various open-source language models perform when working with Ghanaian languages. We have evaluated the performance of over 14 open-source and closed models across 43 languages, using at least 300 sentences per language. Paper publication coming soon.

Why Nsanku Matters

As AI engineers work to bring Ghanaian languages into large language models, it’s essential to have reliable evidence on how existing models perform. Nsanku provides insights that help developers identify which LLMs are most suitable for use in translating content in Ghanaian languages, and which languages currently have stronger or weaker support.

Results (Overview)

Language Specific Results

Detailed performance breakdown for each of the 43 evaluated languages.

Click to expand language-specific charts
Abron
Gikyode
Dangme
Siwu
Anyin
Avatime
Bisa
Bimoba
Southern Birifor
Tuwuli
Ntcham
Buli
Anufo
Dagbani
Southern Dagaare
Ewe
Fante
Ga
Gonja
Farefare
Hanga
Konni
Kusaal
Lelemi
Sekpele
Mampruli
Deg
Nawuri
Chumburung
Nkonya
Delo
Nyagbo
Nzema
Esahie
Paasaal
Tumulung Sisaala
Selee
Tafi
Tampulma
Twi
Vagla
Konkomba
Kasem

Contributing

We welcome contributions from the community! To contribute:

  1. Run the evaluation using our Google Colab notebook
  2. Share your results with us
  3. We'll include your findings in our collective results

Get started with the evaluation notebook:

Evaluated Models

We are running evaluations of these models:

  • gpt-4.1
  • claude-sonnet-4-5
  • gemini-2.5-flash
  • deepseek-v3.1
  • gemma-2-9b-it
  • gemma-2-27b-it
  • gpt-oss-120b
  • kimi-k2-instruct-0905
  • llama-3.1-405b-instruct
  • llama-3.3-70b-instruct
  • llama-4-maverick-17b-128e-instruct
  • mistral-medium-3-instruct
  • qwen3-235b-a22b
  • qwq-32b
  • seed-oss-36b-instruct

Languages Evaluated

The project evaluated 43 Ghanaian languages:

Language Language Language Language
Abron Gikyode Dangme Siwu
Anyin Avatime Bisa Bimoba
Southern Birifor Tuwuli Ntcham Buli
Anufo Dagbani Southern Dagaare Ewe
Fante Ga Gonja Farefare
Hanga Konni Kusaal Lelemi
Sekpele Mampruli Deg Nawuri
Chumburung Nkonya Delo Nyagbo
Nzema Esahie Paasaal Tumulung Sisaala
Selee Tafi Tampulma Twi
Vagla Konkomba Kasem

Contributors

Thanks to our awesome contributors who made it possible:

  1. Onesimus Addo Appiah
  2. Mich-Seth Owusu
  3. Jonathan Asiamah
  4. Elias Dzobo
  5. Kelvin Newman
  6. Edmund O. Benefo
  7. Gerhardt Datsomor
  8. John Ayernor

Contact

For questions or comments, please email natural.language.processing.gh@gmail.com.

To submit your contributions, send them to michsethowusu@gmail.com.

License

This is an open community project. We welcome researchers, developers, and language enthusiasts to participate and help advance NLP for Ghanaian languages.

About

Evaluating zero-shot translation performance of LLMs for Ghanaian languages

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •  

Languages