MiniMax: Root Cause Analysis of Why Large Models Cannot Output the Name "Ma Jiaqi"

robot
Abstract generation in progress

Crypto界消息,MiniMax发布技术博客,披露其m2系列大型模型无法输出人名「马嘉祺」的根因排查过程。排查从一个个例出发,最终揭示了一个影响整个词表的系统性退化问题。根本原因是分词器在训练时将「嘉祺」合并成了一个独立的token。预训练阶段模型见过大量互联网文本,学会了这个token但后续训练的对话数据中,包含「嘉祺」的样本不到5条。后续训练过程中,tool_call标记、代码符号等高频token持续更新周围的向量空间,把「嘉祺」这类低频token挤到了错误的方向。模型仍然「认识」马嘉祺,能准确回答相关信息,丢失的只是输出这个token的能力。团队随后对约200ktoken的完整词表做了全量扫描,发现约4.9%的token发生了显著退化。退化最严重的是日语:29.7%的日语token显著退化,远超韩语3.3%、俄语3.7%、中文3.9%和英文3.5%。

View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • Comment
  • Repost
  • Share
Comment
Add a comment
Add a comment
No comments
  • Pin