Unverified50% confidenceFactExact time
该模型训练分词器得到的词表大小为3266
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
Unverified叠加cv=5的5折交叉验证后，144种参数组合需训练144×5=720次模型58% similarUnverifiedMIT 6.390《机器学习导论》的前身为6.03657% similarVerifiedBF16相比FP16拥有与FP32相同的8位指数位宽，因此动态范围更大，已成为当前主流大模型训练精度选择56% similarUnverifiedLLaMA 2的32000词表在4096维隐层下嵌入层需占用约1.3亿参数56% similarUnverified根据Chinchilla缩放定律，一个30B模型的最优训练数据量约为600B-1T tokens56% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/827401API
curl https://kongchang.com/api/v1/knowledge/claims/827401MCP
get_claim(id=827401)