Feature #2044
Fires on “ read” (not labeled yet: read the examples)
- Fires on
- 0.42% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 90%
- story 7%
- chat 2%
- python 1%
- worksheet 0%
Makes the model say
+: is >0r9.
Makes it not say
there moral repeat said fruit had listened it
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.76<|bos|><|system_start|>tools: edit, glob, grep, question, read, shell, sk
-
agent 3.7<|bos|><|system_start|>tools: edit, glob, grep, question, read, shell, sk
-
agent 3.68<|bos|><|system_start|>tools: edit, glob, grep, question, read, shell, sk
-
agent 3.65<|bos|><|system_start|>tools: edit, glob, grep, question, read, shell, sk
-
agent 3.62<|bos|><|system_start|>tools: edit, glob, grep, question, read, shell, sk
-
agent 3.6<|bos|><|system_start|>tools: edit, glob, grep, question, read, shell, sk
-
agent 3.6<|bos|><|system_start|>tools: edit, glob, grep, question, read, shell, sk
-
agent 3.59<|bos|><|system_start|>tools: edit, glob, grep, question, read, shell, sk
-
agent 3.58<|bos|><|system_start|>tools: edit, glob, grep, question, read, shell, sk
-
agent 3.57<|bos|><|system_start|>tools: edit, glob, grep, question, read, shell, sk
-
agent 3.57<|bos|><|system_start|>tools: edit, glob, grep, question, read, shell, sk
-
agent 3.57<|bos|><|system_start|>tools: edit, glob, grep, question, read, shell, sk