You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Reproduction pack for "Auditing Bias Detection Under Source Shift" (AACL-IJCNLP 2026, main conference). 10,304 audited sentences from three sources with counterfactual pairs, a two-phase LLM audit validated against human annotation, and a pair-aware, leave-one-source-out evaluation protocol.
Modelos de linguagem em português repetem estereótipos? CrowS-Pairs em PT (360 pares), viés em 3 modelos e ablação de cabeças de atenção. PIC/CEUB 2025–2026.
Déjà vu Bias: your bias benchmark has been here before. CrowS-Pairs and StereoSet leaked into OLMo-1.7's pretraining data through the instruction-tuning supply chain (SuperNI → FLAN → Tulu → Dolma), up to 1,728 copies per sentence. We trace the leak into the training shards and measure its effect on bias scores: none detectable.