Feature #2262
Fires on “<|arg|>” (not labeled yet: read the examples)
- Fires on
- 1.04% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 87%
- chat 10%
- story 2%
- python 1%
- worksheet 0%
Makes the model say
pathexpressionoldStringreplaceAllnewString<|arg|>commandSet
Makes it not say
one4 is His and` that Tim
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 7.27it"<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=hi
-
agent 7.23.<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=tools.
-
agent 7.06that returns a to the power of b.<|user_end|><|assistant_start|><|tool_call_start|>write<|arg|>path=math_
-
chat 7.05<|assistant_end|><|user_start|>Now add 7 to the result.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=65
-
chat 7.05<|assistant_end|><|user_start|>Now add 6 to the result.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=95
-
chat 7.04<|assistant_end|><|user_start|>Now add 1 to the result.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=49
-
chat 7.04<|assistant_end|><|user_start|>Now add 1 to the result.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=49
-
chat 7.03<|user_start|>Now add 41 to the result.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=18
-
chat 6.99.<|assistant_end|><|user_start|>Add 61 to it.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=10
-
chat 6.87of 24 and 1.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=24
-
chat 6.8583 and 261 for me?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=18
-
chat 6.85<|user_start|>Now add 79 to the result.<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=56