Feature #4078
Fires on ālā (not labeled yet: read the examples)
- Fires on
- 0.10% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 49%
- story 31%
- worksheet 12%
- chat 7%
- python 1%
Makes the model say
their his a her smiled unique them the
Makes it not say
function hurtsscript codeWrite checksICENSEā
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 1.79Fix the failing test.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 1.69Do the tests pass?"<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 1.66are in this project?<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/*.
-
agent 1.5agent harness.<|system_end|><|user_start|>Fix the failing test.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 1.48files<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/*.
-
agent 1.46.<|system_end|><|user_start|>Find all .py files.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/*.
-
agent 1.46<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 1.46<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 1.45harness.<|system_end|><|user_start|>Check the tests.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 1.44harness.<|system_end|><|user_start|>Check the tests.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 1.44harness.<|system_end|><|user_start|>fix the failing test.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 1.43name.<|assistant_end|><|user_start|>Check the tests.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/