Skip to main content
Back to catalog

$PostalCodesNL

A Dutch postal-code API fragment. The signature under-trained token of the cl100k lineage, “migrating” into Llama-3 and Qwen vocabularies via shared BPE merges.

Raw token

"$PostalCodesNL"

Observed behaviors

  • 01Rank #1 most under-trained in the Magikarp verification for Llama-3-8B/3.1-8B: input-embedding L2 norm ≈1.6e-21, repetition-verification max_prob ≈4.6e-05.GitHub
  • 02It also heads the under-trained lists of Llama-3-70B and the whole Qwen family — a shared under-trained token of the cl100k-lineage tokenizers.arXiv
Observed or documented on
Llama 3Qwen
Behavior types
Under-trained
Tokenizer
Llama-3 BPE / Qwen2Tokenizer
Discovered by
Sander Land & Max Bartolo

References

Related tokens

$PostalCodesNL\tTokenNameIdentifier