Back to catalog
␣サーティ
The katakana 「サーティ」 (“thirty”), traced to a Puzzle & Dragons × Baskin-Robbins (“31 Ice Cream” in Japan) collaboration character name.
Raw token
" サーティ"
␣marks a real space; the symbol itself is not part of the token, and copying preserves it.
Observed behaviors
- 01ChatGPT failed only on the katakana for “thirty/thirty-one” (サーティ/サーティワン) — every other numeral in katakana worked fine.LessWrong
- Observed or documented on
- GPT-3
- Behavior types
- Unspeakable
- Tokenizer
r50k_base- Discovered by
- Jessica Rumbelow & Matthew Watkins