Skip to content
mini-lab
Log in Sign up
Features of prelude-2

Feature #3584

Fires on “ They” (not labeled yet: read the examples)

Fires on
0.34% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 40%
  • story 38%
  • python 12%
  • chat 9%
  • worksheet 1%

Makes the model say

emo ideaexcar!")<|think_start|><|system_end|> True

Makes it not say

there3 itw Itill8 we

Strongest examples

On text the dictionary never trained on. The underlined token is where it fires most.

  1. story 1.67
    rake away and started to play. They both had a fun
  2. story 1.58
    bikes around the bush and played a game of tag. They laughed and had fun
  3. agent 1.56
    AI agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>"how does is_even
  4. story 1.53
    ⏎Tim said, "Okay!" They started to play the game. They had to find two
  5. agent 1.52
    AI agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>Show me listlib.py
  6. agent 1.51
    OpenCode, a coding agent harness.<|system_end|><|user_start|>List all files.
  7. chat 1.51
    Lily and Tom began to repeat the same steps over and over. They laughed and had so
  8. chat 1.5
    want to dance with me?" Tom nodded and they started to dance. They held each other's
  9. story 1.49
    In the end, Tim won the race, and Max learned not to tease others.<|bos|>
  10. story 1.48
    with the window. From then on, Jack knew that he could always support his family
  11. story 1.48
    't want medicine. It is yucky. It makes me suffer." She
  12. agent 1.47
    AI agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>Show me tools.py