supserrr/wdyri-afroxlmr

WDYRI: Swahili mobile-money scam SMS classifier

Davlan/afro-xlmr-base fine-tuned to flag Tanzanian Swahili SMS that try to steal money ("ni tumie kwa namba hii" scams, fake prizes, fake agents, landlord impersonation). Trained with number-balanced counterfactual edits so that the presence of a phone number is not, by itself, evidence of a scam (experiment E11 in the repository).

• Labels: 0 = not scam, 1 = scam. Use the decision threshold 0.5 to match the reported results: the highest threshold up to 0.5 that still catches at least 95% of validation scams.
• Input must go through the project's preprocess() (phone numbers, amounts and links are masked).
• Training data: the BongoScam dataset (MIT licence), template-disjoint split, 760 messages.
• Known blind spots: impersonation scams with no money words (landlord "this is my new number"), which the companion character n-gram model catches; Chichewa and other languages. See the repository for the full evaluation.

Code and results: https://github.com/supserrr/wdyri