Large Language Model
A social ‘study buddy’ gets a conversational lift from Meta Llama
June 6, 2024
6 minute read

FoondaMate—or “study buddy” in Zulu—is a fast-growing, always-on, and highly relatable study aid for middle and high school students in emerging markets that’s built with Meta Llama. Students can ask their AI-enabled study buddy questions on WhatsApp and Messenger and receive conversational replies that help them with their schoolwork.

Many of the 3 million students who use FoondaMate sit in classes of 50 to 70 students, where individual teacher attention is scarce and a textbook for each child isn’t guaranteed. These teenagers come to FoondaMate for help as they learn English, seek to clarify challenging concepts, download public-domain study materials, or prepare for exams that will determine their university eligibility.

FoondaMate founders Dacod Magagula and Tao Boyle have used Meta Llama 2 ever since its first release to enhance conversation flow and tone—key ingredients to the bot’s success. Llama 2 has helped the bot know how to rephrase content for different levels of English comprehension, when to include local language and slang, and even which emojis to add and where. The team has begun testing Meta Llama 3 and is especially excited about the improvement in reasoning capabilities and how it will help enhance multi-step guidance for students.

“The impact of a 24/7 study buddy, accessible on an application students already use, is utterly transformative in environments like this,” says Boyle. “LLM (large language model) technology, when paired with a localized knowledge base and delivered in a way that is affordable and easy to use, may become the single most transformative tool for equal access to education we have ever seen.”

For a bot, being nimble is critical to conversation

FoondaMate’s founding was inspired by Magagula’s experience as a young learner in under-resourced schools, where he often didn’t have access to textbooks. He used the internet to download study papers and notes to advance his education, earning him high marks and a top spot at one of Africa’s most prestigious universities. Now, other students can replicate Magagula’s journey by adding FoondaMate as a friend on WhatsApp and Messenger.

Students can access FoondaMate on WhatsApp and Messenger to ask questions and advance their studies.

Since launching in 2020, FoondaMate has accumulated vast insight into how teenagers in emerging markets chat, learn, and engage. Its largest markets are South Africa, Zimbabwe, and Nigeria. To date, students have asked more than 100 million questions. Boyle says FoondaMate’s impact includes meaningful improvements in students’ grades—with a 30% uplift in university eligibility among students in South Africa who use FoondaMate, compared to their peers who don’t. “As an open-source tool, Llama is the perfect model for training on tone and deploying at scale to resource-constrained students who might struggle to pay more than $2 a month for a study tool,” Boyle says.

A conversation that flows naturally and mimics a teen’s way of texting with friends is core to delivering FoondaMate’s educational materials in an accessible and engaging way. This is especially helpful in Africa, where people often speak multiple languages in one conversation. Additionally, because most of FoondaMate’s users are teenagers, they tend to use slang specific to their age group, sprinkle their texts with lots of emojis, and phrase things in ways that might initially seem counterintuitive.

Taking care with trust and safety

Boyle and Magagula say that they introduced Llama into FoondaMate incrementally—first deploying it in limited contexts, and then gradually expanding its use as their training set grew.

With Llama as a core component of Foondamate, the founders say they appreciate being part of a global open-source community where additional tools and learnings are shared.

“We’re especially excited by the strong focus on trust and safety with regards to human interaction,” Boyle says. “We certainly see ourselves using an increasing number of the safeguard mechanisms and tools being created by the Llama community.”

The team started with a small training set to infuse personality into conversations with FoondaMate, using Llama for non-educational conversations—for example, where users ask “How are you?” or “How was your day?” This allowed them to use generative AI in interactions without any risk to the educational context.

Over time, the team expanded its use in a conversational context, layering Llama 2 on top of FoondaMate’s own education knowledge base and models. With Llama 2 now fine tuned on the tone and personality of the FoondaMate bot, it can rephrase output from FoondaMate’s knowledge base and custom models—even using emojis in the proper context.

Getting excited about learning with FoondaMate

For many of the students who message with FoondaMate, this is likely the first time they’ve ever interacted with technology that feels as though it has been built specifically for them. Boyle adds that the majority of students who use FoondaMate do not have access to computers at home, and some don’t have computers at school. As a result, messaging FoondaMate is usually one of their first encounters with technology outside of interacting with friends on social media and messaging applications. The FoondaMate team wanted to make sure the onboarding process felt as smooth and natural as possible.

FoondaMate's founders said it was important to make sure the study tool was easy to use and accessible.

“We fine-tuned Llama with the goal of keeping our users engaged and feeling as though they are speaking to their smartest friend and not a robotic computer system,” Boyle says.

That approach is paying off, with students responding exactly as the FoondaMate creators had hoped.

“We see their questions shift from ‘How old are you?’ and ‘Where did you go to school?’ to questions about how robots learn and how to make their own robot like FoondaMate,” Boyle says. “It’s been incredibly rewarding to see them getting excited about technology and what’s happening under the hood. Our dream when starting FoondaMate was to help anyone from anywhere get access to a great education and get excited about learning. It’s amazing to see that start to happen.”


Our latest updates delivered to your inbox

Subscribe to our newsletter to keep up with Meta AI news, events, research breakthroughs, and more.

Join us in the pursuit of what’s possible with AI.

Related Posts
Computer Vision
Introducing Segment Anything: Working toward the first foundation model for image segmentation
April 5, 2023
MultiRay: Optimizing efficiency for large-scale AI models
November 18, 2022
ML Applications
MuAViC: The first audio-video speech translation benchmark
March 8, 2023