Skip to main content
Back to catalog

ÃÂÃÂÃÂÃÂ

Classic mojibake token: UTF-8-misdecoded Ã/Â sequences repeated, merged into a single BPE token because such mis-encoded text was frequent in the tokenizer corpus but absent from model training. A 16-repetition variant and `ÛÛ` belong to the same family.

Raw token

"ÃÂÃÂÃÂÃÂ"

Observed behaviors

  • 01Member of the original 140-token anomalous set exhibiting the “unspeakable” failure mode in ChatGPT/GPT-3: refusals, evasions, non-sequitur completions.LessWrong
Observed or documented on
GPT-2GPT-3
Behavior types
Unspeakable
Tokenizer
r50k_base
Discovered by
Jessica Rumbelow & Matthew Watkins

References