Feature #1950
Fires on “ Thank” (not labeled yet: read the examples)
- Fires on
- 0.20% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 45%
- story 33%
- chat 13%
- python 6%
- worksheet 3%
Makes the model say
t pleasew0way trick timeoo
Makes it not say
They they on oneown#ipor
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.83my picture more interesting and beautiful. I improved my drawing. Thank you, teacher."
-
story 1.77. You teach me good things. I will do better next time. Thank you for helping me
-
story 1.61asked, "Can I buy this candy, please?" His mom
-
story 1.6am a magic bird. I can change into any animal I want. Thank you for talking to
-
story 1.57is fine. See? It can still soar. Let me clean up the juice and
-
story 1.56thanked the boy and said, "That's very nice of you. Thank you for being honest
-
story 1.54will be good. I will share. I will make others happy. Thank you, shoe
-
story 1.54ashamed. He said, "I'm sorry. I was wrong. Thank you for helping me
-
story 1.5bird free.⏎When the bird was free, she said, "Thank you, Fred and
-
story 1.48, Lila said goodbye to the thief. She said, "Thanks for being my
-
story 1.45, I understand. Thank you for forgi
-
story 1.42She asked her mom, "Can I be in the parade, please?"⏎Her