Jacaranda launches open source LLM in five African languages

16.08.2024

Last week, we expanded UlizaLlama (AskLlama), our open-source Large Language Model (LLM), to provide AI-driven support in multiple African languages, including Swahili, Hausa, Yoruba, Xhosa, and Zulu. The new multi-lingual model will help deepen how we support new and expectant mothers at scale, while carving new in-roads for AI-driven services across Africa in other sectors.

How does a multi-lingual LLM support new and expecting mothers across Africa?

Off-the-shelf Large Language Models, or LLMs, are typically ineffective in low-resource settings, in part because they’re not adapted to work in languages with limited training data, or customized to specific ‘domains’, like health, agriculture, or education.

In October 2023, we developed the world’s first Swahili-speaking LLM to address this challenge. Our technology team extended the capabilities of Meta’s Llama2, trained the model to respond to general Swahili queries, and then customized it to work within our use case – personalized mHealth support for Kenyan mothers.

In July 2024, we extended this model to Hausa, Yoruba, Xhosa, and Zulu, to reflect our scale ambitions for PROMPTS into Nigeria and South Africa, and as a stepping stone towards our broader ambition of reaching all mums with lifesaving information. Our tech team accomplished this by replicating the process used in the Swahili LLM development: pre-training Meta’s Llama3 for each language, merging the pre-trained models, and finetuning the combined model to create multiple Multilingual LLMs.

We saw promising results in medical accuracy, fluency, and contextual coherence – and we have subsequently integrated the model into our digital health platform, PROMPTS.

Figure 1. To evaluate the quality of generated responses, we extracted 2,500 questions and the corresponding responses from our repository of queries. Our UlizaMama models outperformed the base models across all metrics, including exact matches, semantic similarity, correspondence, and fluency, as measured by AutoEval-based metrics (BertScore, BLEURT, METEOR, BLEU).

We will continue to monitor how these models perform, make improvements as needed, and look to add additional languages, like Twi as we scale our PROMPTS pilot in Ghana. Given the sensitive nature of the information we handle, we will continue to keep humans-in-the-loop for all LLM-related functions, including having responses reviewed and edited before they are sent out to mothers.

Example question and response in Yoruba, showing the decisions the LLM makes to define language, ‘intent’ (message category), accuracy of categorisation, and urgency.

Why does this matter to teams across Africa?

We have made UlizaLlama publicly available, allowing teams with limited resources and budgets to easily ‘plug and play’ the model on their own servers. This setup ensures that teams maintain complete control over their data (which is critical for any sensitive information, like health data), unlike off-the-shelf models such as ChatGPT.

Our hope is that these different language models will be used to extend AI-driven support into other sectors beyond MNH. Multilingual LLMs could assist non-profit organizations in South Africa by tailoring health educational materials in Xhosa or Zulu, aid agriculture development by empowering Nigerian farmers to increase productivity and resilience to climate change (sharing agricultural advice, weather forecasts in Yoruba), or enable researchers to collaborate more effectively by translating complex scientific papers to and from Hausa.

Interested innovators can access and use any of the five language models, including a combined multi-lingual model, via our Model Cards below.
Swahili / English Llama3 Model
Hausa / English Llama3 Model
Yoruba / English Llama3 Model
Xhosa / Zulu / English Llama3 Model
Multilingual Model (Swahili / Hausa / Yoruba / Xhosa / Zulu / English) Llama3 Model