Feature #703
Fires on “re” (not labeled yet: read the examples)
- Fires on
- 0.75% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 76%
- chat 14%
- story 7%
- worksheet 1%
- python 1%
Makes the model say
patternppy[::d dog miniMIT
Makes it not say
shiny shapes knew working called next told heavy
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.79.<|system_end|><|user_start|>How does how_many work?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=
-
agent 3.77case.<|assistant_end|><|user_start|>Where can I find shout?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=
-
agent 3.77_length everywhere.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=
-
agent 3.75<|system_end|><|user_start|>Rename double_all to doubled everywhere.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=
-
agent 3.67to first_letter<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=
-
agent 3.67x)`.<|assistant_end|><|user_start|>Set LIMIT to 7<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=
-
agent 3.67rename the function negate to new_negate"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=
-
agent 3.66[8, 6, 2].<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=
-
agent 3.64harness.<|system_end|><|user_start|>"How does smallest work?"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=
-
agent 3.63harness.<|system_end|><|user_start|>Explain the function hello.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=
-
agent 3.63all the function new_excite instead of excite.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=
-
agent 3.63<|user_start|>Call the function add instead of plus.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=