Feature #1368
Fires on โ,โ (not labeled yet: read the examples)
- Fires on
- 0.58% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 86%
- story 8%
- python 2%
- worksheet 2%
- chat 2%
Makes the model say
day me nowdef then yourgy
Makes it not say
echorepeat repeatared shelfm homesnd
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 5.56<|bos|><|system_start|>tools: edit, glob, grep, question
-
agent 5.56<|bos|><|system_start|>tools: edit, glob, grep, question
-
agent 5.55<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question
-
agent 5.55<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question
-
agent 5.55<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question
-
agent 5.55<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question
-
agent 5.55<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question
-
agent 5.49325<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question
-
agent 5.48(s)`.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question
-
agent 5.4723<|bos|><|system_start|>tools: edit, glob, grep, question
-
story 5.39saw some blood. She started to cry.โ"Ow, ow,
-
agent 5.39_case(s)`.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep, question