Skip to main content
Back to catalog

\uEFC0

A Unicode Private Use Area character that became a single Mistral-family token; it heads the verified under-trained list for nearly every Mistral-tokenizer model.

Raw token

""

Observed behaviors

  • 01Verified under-trained in Magikarp repetition verification across the whole Mistral family (Mistral-7B v0.1–v0.3, Mixtral-8x7B, and derivatives such as Zephyr-7B) — models fail to reproduce it.GitHub
Observed or documented on
Mistral
Behavior types
Under-trained
Tokenizer
Mistral SentencePiece BPE
Discovered by
Sander Land & Max Bartolo

References