#29Hacker News2026-07-04 10:34:02 Dispersion loss counteracts embedding condensation in small language models 分散损失可防止小语言模型嵌入过度凝聚 查看原文 →