工程记录
该分类下共有 13 篇文章
返回所有文章 →Decoding Strategies
Decoding strategies are the decision rules that turn a model’s next-token scores into actual tokens. The model provides a distribution over the vocabulary; the decoder decides how to pick the next token, how to stop, and how to trade off diversity vs. reliability.
1387 个字词
|
7 分钟

