Feature #440
Fires on “<|arg|>” (not labeled yet: read the examples)
- Fires on
- 0.95% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 84%
- chat 7%
- story 5%
- python 3%
- worksheet 1%
Makes the model say
patternpcommanditecase way laptopllo
Makes it not say
⏎⏎⏎. - wore2 keepWow
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.67<|system_end|><|user_start|>"What does quiet do?"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def quiet
-
agent 2.66<|user_start|>"Where can I find power?"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def power
-
agent 2.64<|assistant_end|><|user_start|>Where can I find shout?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def shout
-
agent 2.63<|system_end|><|user_start|>"What does odd return?"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def odd
-
agent 2.63<|user_start|>"Where can I find double?"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def double
-
agent 2.6<|system_end|><|user_start|>Where can I find biggest?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def biggest
-
agent 2.59Where can I find max_item?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def max
-
agent 2.59<|system_end|><|user_start|>Where can I find exclaim?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def exclaim
-
agent 2.58<|system_end|><|user_start|>Where can I find contains?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def contains
-
agent 2.55<|system_end|><|user_start|>Where can I find tail?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def tail
-
agent 2.55<|system_end|><|user_start|>"What does backwards return?"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def backwards
-
agent 2.55where can I find last_letter?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def last