Hacker News Reader

No mention of calibration. Is it just another llm finetune?

> Our post-training utilizes label-smoothed cross-entropy for valid schema outputs paired with a Brier loss to refine probability calibration.