System Architecture: SmolLM2-135M (Fine-tuned via PEFT/LoRA)
Operational Target: 2GB RAM Mobile SoC (Low-Power ARMv8)
🏗️ 1. Technical Abstract
Smol-AI-Africa represents a breakthrough in Low-Resource Natural Language Processing (LR-NLP). While modern LLMs are scaled toward trillion-parameter architectures, this project focuses on Extreme Optimization for the African digital frontier.
We avoid full-parameter updates to prevent Catastrophic Forgetting. We apply a low-rank decomposition to the weight updates:
$$W = W_0 + \Delta W = W_0 + BA$$
Using a Rank (r) of 16 and Alpha of 32, we target the q_proj and v_proj modules for maximum efficiency on 2GB RAM devices.
🌍 3. Socio-Economic Impact
3.1 Mobile Accessibility
In Kano—Africa's commercial nerve center—hardware is the gatekeeper. By optimizing for 2GB RAM, we ensure this AI runs on second-hand smartphones and older Android devices used by the 'Street Smart' hustle.
