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

Feature #2168

Fires on “1” (not labeled yet: read the examples)

Fires on
0.50% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 52%
  • chat 23%
  • worksheet 20%
  • story 4%
  • python 1%

Makes the model say

* typoSet utils fast itsrun

Makes it not say

." girl little new have<|bos|> big⏎

Strongest examples

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

  1. chat 4.02
    <|user_end|><|assistant_start|><|think_start|>65+08: 5+8=13, 3
  2. chat 4.01
    now?<|user_end|><|assistant_start|><|think_start|>8+5: 8+5=13, 1
  3. chat 3.91
    <|user_end|><|assistant_start|><|think_start|>65+08: 5+8=13, 3
  4. chat 3.85
    <|user_end|><|assistant_start|><|think_start|>18+85: 8+5=13, 3
  5. worksheet 3.79
    now?⏎08+85: 8+5=13, 3
  6. chat 3.75
    <|think_start|>376+007: 6+7=13, 3
  7. chat 3.75
    <|think_start|>538+005: 8+5=13, 3
  8. chat 3.72
    <|think_start|>008+835: 8+5=13, 3
  9. chat 3.7
    5?<|user_end|><|assistant_start|><|think_start|>8+5: 8+5=13, 1
  10. chat 3.67
    <|think_start|>005+798: 5+8=13, 3
  11. chat 3.66
    <|user_end|><|assistant_start|><|think_start|>05+58: 5+8=13, 3
  12. worksheet 3.65
    007+7786: 7+6=13, 3