Towards Global South AI Sovereignty: A Federated Learning Framework for Collaborative LLM Development
Phemelo Maile, Mahlomola Mohlomi , Mueletshedzi Moses Mubvafhi
This project explores federated fine-tuning as a pathway for collaborative AI development in the Global South. Using African-language sentiment classification as a case study, we simulate five language-specific nodes that train locally on AfriSenti data and share only model updates rather than raw text. The results show that parameter-efficient updates reduce communication cost by about 184×, and that adding more participating nodes improves mean validation accuracy before plateauing. However, the lowest-resource language remains poorly protected, showing that future systems need stronger privacy-preserving and language-aware aggregation methods.
Great work.!
Excellent framing of the dependency risk faced by Global South institutions. Using parameter-efficient updates (LoRA) to reduce communication overhead by approximately 184x is a highly practical approach for resource-constrained environments. To push this to the next level, incorporating secure aggregation or differential privacy will be necessary, as model updates can still leak information under certain attacks. Additionally, exploring personalized adapters could help address the persistent low-resource node performance gap seen with the Xitsonga language node.
A well-written and candid proof-of-concept for an important but currently under-explored challenge: collaborative model development using data-local models in the Global South. It identifies its own limitations clearly. Rather than overstating the benefits of q-FFL, the authors are upfront that it fails to protect the least-resourced node and that simply increasing participation does not resolve performance issues caused by resource constraints. That negative finding is genuinely useful.
Possible ways to improve it:
(1) Provide baseline numbers. For the three-class sentiment task, random or majority is roughly 0.33, while Xitsonga sits around 0.17, which is below chance, so it is unclear whether the federation learned anything about the lowest-resource language. Reporting a random or majority baseline and a centralized (non-federated) upper bound would put context around every result.
(2) Demonstrate statistical significance. Many reported effects (q-FFL's gains of roughly 0.01 to 0.02, and the worst-language trend from about 0.16 to 0.18) fall within the seed-to-seed variance. Paired t-tests or confidence intervals across the three seeds would separate real effects from measurement noise. Figure 2 also looks non-monotonic, so "a modest improvement" may be reading a pattern into noise.
(3) Refine the scope of the framing. State up front that this is an encoder-classification proxy, and treat large-scale generative or instruction tuning as future work (which you already mention).
(4) Present the full efficiency trade-off. The roughly 184x reduction in transmission size is mostly a consequence of sending only the LoRA adapters, so reporting any accuracy cost relative to sending full model updates would show how efficient the system truly is.
(5) Run a sensitivity analysis for q-FFL. You only report q = 2.0. A short sweep over different q values would show whether fairness weighting can ever help the worst-performing node, or whether the failure is inherent to this formulation.
Net: a clear, honest submission with a genuinely useful negative result; the main path to a stronger version is better baselines and significance testing so the currently small effects can be trusted.
Cite this work
@misc {
title={
(HckPrj) Towards Global South AI Sovereignty: A Federated Learning Framework for Collaborative LLM Development
},
author={
Phemelo Maile, Mahlomola Mohlomi , Mueletshedzi Moses Mubvafhi
},
date={
},
organization={Apart Research},
note={Research submission to the research sprint hosted by Apart.},
howpublished={https://apartresearch.com}
}
