Back to catalog
useRal
A C#-flavored identifier fragment (cl100k id 89471) and head of the useRal triplet (useRal/useRalative/useRalativeImagePath) — Watkins flagged the triplet phenomenon itself as worth investigating. Also verified as under-trained in OLMo 2 by Magikarp.
Raw token
"useRal"
Observed behaviors
- 01An “unspeakable” glitch token: ChatGPT often fails to repeat it (blank messages or mid-answer termination); nondeterministic at temperature 0 (bold-flagged in the source).LessWrong
- 02Rank #3 most under-trained in the OLMo 2 Magikarp repetition verification: a max self-repeat probability of just 2.1e-11.GitHub
- Observed or documented on
- GPT-3.5 / GPT-4OLMo 2 (AI2)
- Behavior types
- UnspeakableUnder-trained
- Tokenizer
cl100k_base / GPT2Tokenizer (OLMo 2)- Discovered by
- Matthew Watkins (cl100k); Sander Land & Max Bartolo (OLMo 2)