Feature #121
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 0.45% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 73%
- story 14%
- chat 9%
- python 3%
- worksheet 2%
Makes the model say
theway rest numsuallyva
Makes it not say
plus upon than love hurts explore goneate
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 4.43.py<|arg|>oldString=positive<|arg|>newString=is_positive<|arg|>replaceAll=true<|tool_call_end|>
-
agent 4.42calc.py<|arg|>oldString=absolute<|arg|>newString=magnitude<|arg|>replaceAll=true<|tool_call_end|>
-
agent 4.38.py<|arg|>oldString=total<|arg|>newString=new_total<|arg|>replaceAll=true<|tool_call_end|>
-
agent 4.37<|arg|>oldString=welcome<|arg|>newString=greet<|arg|>replaceAll=true<|tool_call_end|>
-
agent 4.29=first_letter<|arg|>newString=new_first_letter<|arg|>replaceAll=true<|tool_call_end|>
-
agent 4.26<|arg|>oldString=only_even<|arg|>newString=keep_even<|arg|>replaceAll=true<|tool_call_end|>
-
agent 4.22=sum_list<|arg|>newString=new_sum_list<|arg|>replaceAll=true<|tool_call_end|>
-
agent 4.18.py<|arg|>oldString=initial<|arg|>newString=first_letter<|arg|>replaceAll=true<|tool_call_end|>
-
agent 4.13.py<|arg|>oldString=max_of<|arg|>newString=larger<|arg|>replaceAll=true<|tool_call_end|>
-
agent 4.12.py<|arg|>oldString=count<|arg|>newString=how_many<|arg|>replaceAll=true<|tool_call_end|>
-
agent 4.11=has_item<|arg|>newString=my_has_item<|arg|>replaceAll=true<|tool_call_end|>
-
agent 4.1newString=my_tail<|arg|>replaceAll=true<|tool_call_end|>