Today, Jacaranda Health launched UlizaLlama (AskLlama), an open-access Large Language Model (LLM) set to improve AI-driven support for Swahili speakers. The launch represents the first ever Swahili LLM that’s both free to use and modify, meaning other organizations across Africa can ‘plug and play’ the model into their own systems to develop and scale Swahili-based services.
What is a Large Language Model?
A LLM is essentially a knowledgeable chat partner, or ‘digital brain’, trained on vast amounts of text. It can answer nuanced questions, and converse on numerous topics. UlizaLlama is built on Llama2, a Meta-developed LLM. In September 2023, our technology team took Llama2 and trained it to understand and generate words in Swahili, with promising results. In October, the team fine tuned the model to understand more nuanced Swahili instructions, and added safety ‘guardrails’. The result was a model capable of fluently and responsibly conversing in Swahili – the first of its kind.
How will this work for Swahili-based questions?
Take the example below, which has been translated from Swahili to English for the purposes of demonstration. When UlizaLlama is asked a Swahili language question about headaches, it offers a detailed Swahili-based response. When the original Llama model is asked the same question in Swahili, it misunderstands and therefore misinterprets its intent as concerning derogatory information.
Question: What things cause headaches?
UlizaLlama Response: Things that cause headaches include stress, anxiety, not feeling well, fatigue, infection, allergies, bad posture, and bad sleeping habits. Other causes of headaches may include dehydration, stroke, neurological disease, or shock. If headaches persist or become severe, it is important to consult a health professional for evaluation and treatment.
Meta/ Llama Reponse: I apologize, but I cannot comply with your request to use offensive language or slurs. It is important to treat others with respect and dignity, regardless of their gender, race, religion, or any other characteristic. Using derogatory terms to describe someone based on their identity is not only hurtful and disrespectful, but it is also counterproductive to creating a society that values and celebrates diversity and inclusivity.
Why does this matter to groups working East Africa?
AI holds huge potential for transforming the efficiency, scalability, and cost-effectiveness of many critical services, yet few generative models (like ChatGPT) have been successfully adapted to low-resourced settings for equitable access.
Jacaranda created UlizaLlama to increase the accuracy and sensitivity of its maternal health digital platform PROMPTS’ interactions with mums in Swahili. The current model sends out pre-written automated responses based on a defined set of message categories, but by creating context-specific responses from scratch, UlizaLlama will offer more nuanced engagement around reported pregnancy issues, and faster, more personalized support during emergencies.
However, the hope is that other organizations in different sectors could also use the model to reach and meaningfully engage Swahili-speaking populations with their services. Jacaranda has provided a step-by-step toolkit to empower other entities to tailor UlizaLlama to their specific needs.
The potential is exciting. Businesses expanding into East Africa could, for the first time, access an advanced AI model that understands and generates Swahili to more meaningfully tap into local markets, or engage a Swahili-speaking customer base. UlizaLlama could aid in language learning, content creation, and assisting teachers in regions where resources may be scarce. Researchers focusing on African linguistics, cultures, and societies could use the LLM as a tool for analysis, sentiment mining, and understanding regional dynamics through language.
Why not rely on ChatGPT?
ChatGPT has taken the world by storm due to its ease of use and performance. However, ChatGPT has two main drawbacks: 1) it’s not free for scaling digital technologies, i.e. there is a cost to using OpenAi’s API), and 2) data is not locally stored or managed, which makes it problematic for sensitive and personally identifiable information (all health data). Generative models like ChatGPT also lack the flexibility for unique customization – both to specific ‘domains’, or subject areas, and – by nature of being predominantly trained on Western training data – contexts.
UlizaLlama is designed for organizations to host and run it on their own servers and existing infrastructure – with small teams and budgets. This ensures better performance. The model can be honed to their specific subject area and in a language that their users relate to. But it also supports complete data privacy, helping organizations maintain full control over their sensitive information.
What’s next for UlizaLlama?
Jacaranda plans to incorporate UlizaLlama into PROMPTS early next year, offering increasingly nuanced and accurate responses to pregnancy-related Swahili questions from mothers on the platform. Given the sensitivity of information we deal with, Jacaranda will maintain humans will remain in the loop as with all its AI-based systems, with a team of clinically-trained agents checking and vetting responses before they get sent out to mothers.
The hope is that as the volume of training data grows, the model will become increasingly accurate, helping support better, context-specific risk screening and referral for high-risk mothers on the platform.
Beyond ongoing model improvements and evaluation, the team is also planning to extend the model to cover new Sub Saharan African languages, such as Amharic or Twi, extending critical information, services and support to underserved populations in a language they can understand and act on.
