The global AI landscape is rapidly evolving, and not always in the right way. Organizations around the world are grappling with the same challenges of data privacy and security, bias and fairness, and regulatory compliance. In low-resourced contexts, additional issues of contextualization, language adaptation, and costly integrations continue to curb the potential impact of this cutting edge technology in areas where it’s needed most.
In July, members of Jacaranda’s quickly-growing technology team, Stanslaus and Sylvia, traveled to Bangalore, India, for a cross-sectoral immersive workshop on Large Language Models, or LLMs, hosted by our partners at The Agency Fund. The three day event aimed to:
- Assess how LLMs were being implemented and scaled across different industries and sectors
- Identify common challenges to implementation
- Facilitate knowledge sharing on new techniques, tools, and advancements in AI and LLM technology
As part of this event, Stanslaus and Sylvia took participants through Jacaranda’s journey with LLMs. Late last year, the team rolled out the world’s first Swahili-speaking LLM, UlizaLlama (AskLlama), built on Meta’s Llama2 model. We have since made this domain-specific (to answer pregnancy-specific questions from mothers on PROMPTS), and available in other ‘low-resource’ languages, including Hausa, Yoruba, Xhosa, Zulu to align with early scoping efforts for expansion into Nigeria and South Africa respectively.
We were able to share lessons from this work with other implementers, specifically around building and fine-tuning LLMs for different use cases, and techniques for combining multiple LLMs (eg. different language variations) into a single model (eg. a multilingual model).
Our team were pleased to meet new innovators in the field and reconnect with some familiar faces, including Karya, Udhayam, IDInsight, Kabakoo Team, and Noora Health. Stanslaus and Sylvia returned to Nairobi with a host of new learnings to help further our LLM work in PROMPTS, including emerging trends in LLM evaluation, prompt engineering techniques, effective implementation of RAG, and ethical considerations for LLMs.
