Jev 校准指标 ECE 反而不如 ModernBERT

Charly Poly @whereischarly原推发布于 2026.09.18

在 X 上打开原推

Now, the part nobody is testing. Jev is trained with RLCD ("Reinforcement Learning for Calibrated Decisions"), calibration is literally in the name. ECE (lower is better): - ModernBERT — 0.081 - Jev — 0.105 The open encoder is better calibrated than the model built for calibration.