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

Feature #2112

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

Fires on
0.37% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 68%
  • story 18%
  • python 9%
  • chat 3%
  • worksheet 2%

Makes the model say

justlower somethingBut littleelyselves But

Makes it not say

w(- can){Yes True than

Strongest examples

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

  1. story 2.09
    . They had a great time eating and laughing together.⏎In the end, Fred's foot
  2. story 2.09
    laughed and shared their sandwiches with the birds.⏎In the end, Tim and his
  3. story 2.08
    in the forest came to watch and admire them.⏎In the end, Blue and R
  4. story 2.07
    games. The rain didn't bother them anymore.⏎In the end, Sue and her
  5. story 2.06
    the new color was fun and could make beautiful things too. In the end, all the colors
  6. story 2.05
    Tom laughed, the more the machine made funny faces.⏎In the end, Tom was happy
  7. story 2.02
    swam, and soon they were near the finish line.⏎In the end, Sam and Fin
  8. story 2.01
    fun like Tom, and her dress still got dirty.⏎In the end, the event was
  9. story 2.01
    sed at each other, trying to get the thread.⏎In the end, Spot won the
  10. story 1.99
    identally turned the radio up too loud. The sound was harsh, just like
  11. agent 1.99
    5: print(fact(1))⏎<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=arith
  12. story 1.99
    kid who took the most steps won a prize.⏎In the end, all the kids