Unverified50% confidenceFactExact time
Mid-training（持续预训练）是介于预训练和微调之间的训练阶段，数据量通常达到数千亿到数万亿token
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026
Sources
Related Claims
Unverified当前主流大模型的训练数据量已达到十几T甚至六十T token69% similarUnverified预训练阶段在海量无标注语料上通过自监督学习完成，计算成本极高、周期通常以月计，而指令微调、RLHF对齐等工程阶段可在数周甚至数天内完成68% similarUnverified主流大模型的训练数据通常有6-18个月的滞后期68% similarUnverified在千亿参数大模型训练中，节点间通信开销可能占总训练时间的30%-50%67% similarUnverified训练过程可拆分为外循环（周/月单位，接收用户反馈）和内循环（小时/天单位，快速爬坡验证）67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/51432API
curl https://kongchang.com/api/v1/knowledge/claims/51432MCP
get_claim(id=51432)