Feature #516
Fires on “<|assistant_end|>” (not labeled yet: read the examples)
- Fires on
- 1.09% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 91%
- story 4%
- worksheet 3%
- chat 1%
- python 1%
Makes the model say
<|tool_end|> <|assistant_end|> theyexc": owner'
Makes it not say
room Inside about there In inside open found
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.03oldString=greet<|arg|>newString=hello<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited demo.
-
agent 1.98oldString=hello<|arg|>newString=greet<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited main.
-
agent 1.96oldString=welcome<|arg|>newString=greet<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited test_
-
agent 1.87newString=new_first_letter<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited tools.
-
agent 1.86even<|arg|>newString=keep_even<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited stats.
-
agent 1.85oldString=greet<|arg|>newString=welcome<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited test_
-
agent 1.82oldString=head<|arg|>newString=first<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited test_
-
agent 1.81newString=my_max_item<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited listlib.
-
agent 1.81echo<|arg|>newString=new_echo<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited text.
-
agent 1.8oldString=first<|arg|>newString=head<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited core.
-
agent 1.8oldString=excite<|arg|>newString=exclaim<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited main.
-
agent 1.8total<|arg|>newString=new_total<|arg|>replaceAll=true<|tool_call_end|><|assistant_end|><|tool_start|>Edited test_