Feature #2980
Fires on “r” (not labeled yet: read the examples)
- Fires on
- 0.21% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 56%
- story 27%
- chat 10%
- python 6%
- worksheet 1%
Makes the model say
r huggedueiss longanceight kept
Makes it not say
cat not < overcount man across prints
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 4.57py<|arg|>oldString=absolute<|arg|>newString=magnitude<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 4.56_item<|arg|>newString=my_has_item<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 4.5_list<|arg|>newString=new_sum_list<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 4.45.py<|arg|>oldString=head<|arg|>newString=first<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 4.43_halve<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 4.42even<|arg|>newString=is_even_v2<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 4.38<|arg|>oldString=max_of<|arg|>newString=larger<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 4.34newString=my_tail<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 4.32=sum_list<|arg|>newString=total<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 4.31<|arg|>oldString=positive<|arg|>newString=is_positive<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 4.28<|arg|>oldString=square<|arg|>newString=squared<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 4.23oldString=head<|arg|>newString=head_v2<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>