Feature #646
Fires on “ broke” (not labeled yet: read the examples)
- Fires on
- 0.65% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 82%
- story 12%
- chat 5%
- worksheet 0%
- python 0%
Makes the model say
.sed! something things their into
Makes it not say
function<|user_start|><|assistant_end|>newStringreplaceAll generator exclamation [
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.91But then, something bad happened. The ball hit a rock and it broke. The sy
-
story 3.79put them on, she squeezed them too hard.⏎The sunglasses broke into pieces. Tim
-
story 3.6to play too. The dog jumped and bit the rope. The rope broke, and Tim fell
-
agent 3.55))⏎5: print(is_odd(SIZE))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
story 3.53. The ground shook and people were scared. Some things fell down and broke.⏎One day
-
story 3.52he heard Joe screaming. He had dropped the cell phone and it broke.⏎Steven
-
story 3.48Bad dog, Max!" Ben says, frowning. "You broke our tower! Go
-
story 3.47noises. The smelly hurricane was very angry. It broke the box and came
-
story 3.46got angry and jumped on Tim. The cake fell on the ground and broke into pieces. Tim
-
agent 3.44= k + 1⏎10: return k<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.4310⏎6: print(odd(COUNT))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path
-
agent 3.41]⏎6: print(count_items(NUMBERS))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path