QLoRA fine-tune of
McGill-NLP/AfriqueLlama-8B
on Tanzanian business, tax, company registration, and financial regulation data.
Base model: AfriqueLlama-8B (Llama 3.1 8B, 20 African languages incl. Swahili)
Languages: Swahili (sw), English (en)
Training: QLoRA r=64 on Tesla T4 (sm_75)
Training pairs: 2,537 + batch-013 additions
Validation loss: 0.4447
Gate result: PASSED
Started from: adapter-v7-lora (true continued fine-tuning, r=64 → r=64)
LoRA-only checkpoint: prospAprospA007/africa-giants-adapter-v8-lora
