briankaniaru181-jpg/NIRU_HACKATHON

Sauti AI: Computationally efficient framework for African language conversational AI. Stage 1: Continued pretraining on diverse Swahili text. Stage 2: Conversational fine-tuning. Built with LoRA for accessibility. NIRU AI Hackathon 2026. # Sauti AI A Reproducible, Parameter-Efficient Framework for African Language Conversational AI A two-stage training methodology for culturally grounded Swahili and Kenyan languages ## Executive Summary Sauti AI is a production-oriented conversational AI framework designed for African languages, beginning with Swahili and extending systematically to Kikuyu and other Kenyan languages. The project introduces a validated two-stage training methodology, continued pretraining followed by conversational fine-tuning, implemented using parameter-efficient techniques (LoRA + 4-bit quantization) on accessible hardware. An initial Swahili implementation of this methodology achieved a validation perplexity of 3.97(Stage 2 Fine-tuning), with end-to-end training completed in approximately 9 hours on a single P100 GPU. During the NIRU AI Hackathon 2026, this repository serves both as: 1) A record of validated results, and 2) A reproducible framework, which will be re-executed publicly during the competition period. Because training outcomes depend on stochastic initialization, hardware variance, and data ordering, future runs may produce slightly different metrics. Accordingly, this repository emphasizes methodological reproducibility and engineering rigor, not one-off numerical claims. ### What This Repository Represents A proven training architecture for low-resource African languages A production-quality Swahili MVP, already trained and deployed A live reproducibility exercise, to be re-run during the competition A foundation for Kikuyu expansion, using the same pipeline This distinction is intentional and central to the project’s scientific and engineering integrity. ### Core Methodology: Two-Stage Training ### Stage 1: Continued Pretraining #### Linguistic & Cultural Foundation ##### Objective To strengthen the base model’s internal representation of Swahili grammar, narrative structure, metaphor, and cultural context before any conversational specialization. …