Feature #111
Fires on “l” (not labeled yet: read the examples)
- Fires on
- 0.34% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 74%
- story 20%
- chat 5%
- python 1%
- worksheet 0%
Makes the model say
l theo everyoneO all wasight
Makes it not say
<|user_start|><|system_start|> calculatorva))command function)`.
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 4.34harness.<|system_end|><|user_start|>Check the tests.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 4.25.<|system_end|><|user_start|>"Are the tests passing?"<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 4.2<|system_end|><|user_start|>run the tests and fix any bug.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 4.16<|system_end|><|user_start|>Run the tests and fix any bug.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 4.15.<|system_end|><|user_start|>"Can you run the tests?"<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 4.15<|system_end|><|user_start|>"Show me the Python files."<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/*.
-
agent 4.13agent harness.<|system_end|><|user_start|>Check the tests<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 4.11name.<|assistant_end|><|user_start|>Check the tests.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 4.1<|system_end|><|user_start|>Run the tests and fix any bug.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 4.09<|system_end|><|user_start|>The tests are failing, fix the bug<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 4.05.<|system_end|><|user_start|>"Can you run the tests?"<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/
-
agent 4.04harness.<|system_end|><|user_start|>Check the tests.<|user_end|><|assistant_start|><|tool_call_start|>glob<|arg|>pattern=**/