Feature #3017
Fires on ālā (not labeled yet: read the examples)
- Fires on
- 0.27% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 82%
- python 8%
- chat 7%
- story 2%
- worksheet 0%
Makes the model say
is cheer thebob very sc it
Makes it not say
replaceAll<|think_end|><|assistant_end|><|user_end|><|assistant_start|>ies<|tool_end|>not
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.55.<|system_end|><|user_start|>"Can you run the tests?"<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
chat 2.545`.ā<|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are mini, a small and friendly assistant
-
agent 2.52.<|system_end|><|user_start|>"List the Python files."<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/*.
-
chat 2.522.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are mini, a small and friendly assistant
-
agent 2.5.<|system_end|><|user_start|>"Can you run the tests?"<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
chat 2.49awe.<|bos|><|system_start|>You are mini, a small and friendly assistant
-
chat 2.4710.<|bos|><|system_start|>tools: calculator<|system_end|><|system_start|>You are mini, a small and friendly assistant
-
agent 2.46.<|system_end|><|user_start|>"Can you run the tests?"<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 2.46harness.<|system_end|><|user_start|>Are the tests passing?<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 2.45harness.<|system_end|><|user_start|>Can you run the tests?<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 2.45<|user_start|>The tests fail. Can you fix them?<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 2.45.<|system_end|><|user_start|>Find all .py files.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/*.