Skip to main content
Back to catalog

\\+::\\+

A mysterious markup fragment starting with double backslashes, likely a Chinese-web markup remnant; the #1 most under-trained token in Yi-9B.

Raw token

"\\\\+::\\\\+"

Observed behaviors

  • 01Rank #1 most under-trained in the Yi-9B Magikarp repetition verification (embedding L2 norm 2.12e-06): a max self-repeat probability of just 1.1e-05.GitHub
Observed or documented on
Yi (01.AI)
Behavior types
Under-trained
Tokenizer
LlamaTokenizer
Discovered by
Sander Land & Max Bartolo