Skip to main content
Back to catalog

\u200Cآمباردا

A Persian fragment beginning with a zero-width non-joiner (U+200C ZWNJ, an invisible character); the #2 most under-trained token in Gemma-7B.

Raw token

"‌آمباردا"

Observed behaviors

  • 01Rank #2 most under-trained in the Gemma-7B Magikarp repetition verification (E_out cosine-distance indicator 7.39e-06): a max self-repeat probability of just 4.4e-04.GitHub
Observed or documented on
Gemma (Google)
Behavior types
Under-trained
Tokenizer
GemmaTokenizer
Discovered by
Sander Land & Max Bartolo