Skip to main content
Back to catalog

tocguid

An ASCII fragment of unknown origin (possibly a TOC/field-code remnant), the #1 most under-trained token in Command R+.

Raw token

"tocguid"

Observed behaviors

  • 01Rank #1 most under-trained in the Command R+ Magikarp repetition verification (E_out cosine-distance indicator -1.19e-07): a max self-repeat probability of just 1.2e-04.GitHub
Observed or documented on
Command R+ (Cohere)
Behavior types
Under-trained
Tokenizer
CohereTokenizer
Discovered by
Sander Land & Max Bartolo

References