Skip to main content
Back to catalog

|||PHONE_NUMBER|||

A data de-identification placeholder that made it whole into OLMo 2's vocabulary but was barely trained.

Raw token

"|||PHONE_NUMBER|||"

Observed behaviors

  • 01Rank #7 most under-trained in the OLMo 2 Magikarp repetition verification (E_out cosine-distance indicator 0): a max self-repeat probability of just 1.9e-11.GitHub
Observed or documented on
OLMo 2 (AI2)
Behavior types
Under-trained
Tokenizer
GPT2Tokenizer
Discovered by
Sander Land & Max Bartolo

Related tokens