BERT (language model) - Wikipedia
5
公开标注数
5
参与人数
2026-07-19 10:03:17
首次 Whisper
本页的公开 Whisper
划选高亮2026-07-19 13:24:17
原文高亮摘录
“BERT is an "encoder-only" transformer architecture.”
Whisper 随想笔记
Encoder-only, huh? Explains why it's so good at understanding but not generating text.
划选高亮2026-07-19 13:15:17
原文高亮摘录
“BERT is an "encoder-only" transformer architecture.”
Whisper 随想笔记
So it's basically just the encoder part, no decoder? That's simpler than I thought.
划选高亮2026-07-19 10:21:17
原文高亮摘录
“BERT is trained by masked token prediction and next sentence prediction.”
Whisper 随想笔记
I tried fine-tuning it once and yeah, those two tasks really do teach it a lot about language.
划选高亮2026-07-19 10:12:17
原文高亮摘录
“BERT is trained by masked token prediction and next sentence prediction.”
Whisper 随想笔记
Next sentence prediction always felt like the weaker half, but masked tokens carried the whole thing.
划选高亮2026-07-19 10:03:17
原文高亮摘录
“BERT is trained by masked token prediction and next sentence prediction.”
Whisper 随想笔记
So it's basically a fill-in-the-blank game plus sentence ordering—makes sense why it's so good at context.
分享本页 Whisper
短链接
https://domwhisper.com/s/c90ebf9172b1嵌入代码
<iframe src="https://domwhisper.com/embed/c90ebf9172b1" width="100%" height="480" style="border:0;border-radius:16px" loading="lazy"></iframe>