Feature #3983
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 1.83% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 57%
- python 15%
- worksheet 12%
- chat 12%
- story 4%
Makes the model say
function0⏎ redocletterimase
Makes it not say
missedeven this thereri happy something
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 5.124: acc = acc + 1⏎5:
-
python 5.04if c in "aeiou":⏎ k = k + 1⏎ return k⏎
-
agent 5.03k = k + 1⏎6:
-
chat 4.82<|user_end|><|assistant_start|>One day, a kind cat named Kitty felt a strong urge to go outside
-
agent 4.82acc = acc + 1<|arg|>newString=
-
agent 4.82in "aeiou":⏎5: k = k + 1⏎6:
-
agent 4.69":⏎5: k = k + 1⏎6:
-
agent 4.6in "aeiou":⏎9: k = k + 1⏎10:
-
agent 4.57in "aeiou":⏎ k = k + 1⏎ return k⏎
-
story 4.56with her mom and dad.⏎One day, Emma had an urge to explore the
-
chat 4.38was a girl named Lily. She had a big, pink eraser.<|assistant_end|><|user_start|>
-
chat 4.37was a girl named Lily. She had a big, pink eraser.<|assistant_end|><|user_start|>