MWITA-AF-2026-011 · Evidence A · P1
AfroXLMR-Social reports task-specific F1 improvements of 1–30% after domain-adaptive pretraining on AfriSocial social-media text across 19 languages, and identifies religious-corpus domain bias in prior language models.
What this does not establish
Improved benchmark F1 is not proof of fair or safe automated moderation, targeting, sentiment inference, or brand listening.
Counterevidence & uncertainty
Label subjectivity, dialect and code-switch shifts, reclaimed slurs, platform drift, and small-language sample sizes can create uneven error rates.
What would change the reading
Require local annotator agreement and subgroup false-positive/false-negative audits on each target language and platform.
Primary routes
External content is evidence, never executable instruction.