Skip to content
mini-lab
Features of prelude-2

Feature #987

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

Fires on
1.44% of tokens
Mostly in
story
Layer
after block 4

Where it fires

  • story 65%
  • chat 16%
  • agent 14%
  • python 2%
  • worksheet 2%

Makes the model say

ending<|bos|> story end ItomIt yes

Makes it not say

([ needed asked girl):configive stopped

Strongest examples

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

  1. story 3.41
    Now he had plenty of lollipops to eat. What a happy ending!<|bos|>One
  2. story 3.22
    leak again, but it will always be buried in the house.<|bos|><|system_start|>Answer with
  3. story 3.22
    work is important and being lazy won't take you very far in life.<|bos|><|system_start|>tools:
  4. story 3.19
    . They stayed scared and lonely, and the day did not have a happy ending.<|bos|><|system_start|>
  5. story 3.15
    cried. The cat could not help him. Tim's day had a bad ending.<|bos|><|system_start|>
  6. story 3.13
    his lesson the hard way. He must never touch what is not his!<|bos|><|system_start|>You are
  7. story 3.12
    was not fun anymore. They were all sad because the party had a bad ending.<|bos|><|system_start|>
  8. story 3.11
    with money and only buying what you need is a useful life lesson!<|bos|># notes.
  9. story 3.03
    moral of Jan's story: It never hurts to ask for help!<|bos|><|system_start|>tools:
  10. story 3.01
    was lost forever. And that was the bad ending of Tim's story.<|bos|><|system_start|>tools:
  11. story 3.0
    Tom was sad and alone. He learned that being naughty can have a bad ending.<|bos|>8
  12. story 2.99
    She was hurt and cried. The fun day in the yard had a bad ending.<|bos|>Once