Feature #1462
Fires on “r” (not labeled yet: read the examples)
- Fires on
- 0.42% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 42%
- story 36%
- python 13%
- chat 7%
- worksheet 2%
Makes the model say
ue belong family<|bos|> you happened dis town
Makes it not say
uponreturn whether soon$5kingated
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.98Only a note that said: "You should not take things that are not yours. It
-
story 2.98Ben said no.⏎"Mom said we should not take things that are not ours,"
-
agent 2.96<|arg|>oldString=square<|arg|>newString=squared<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 2.94<|arg|>oldString=welcome<|arg|>newString=greet<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
python 2.88):⏎ if a > b:⏎
-
agent 2.86_item<|arg|>newString=my_has_item<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 2.85_list<|arg|>newString=new_sum_list<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 2.84=sum_list<|arg|>newString=total<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 2.82py<|arg|>oldString=absolute<|arg|>newString=magnitude<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 2.8_letter<|arg|>newString=new_final_letter<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>
-
story 2.8hard lesson today," she said. "You should not take things that are not yours. And
-
agent 2.77new_echo<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>