Feature #1192
Fires on “ we” (not labeled yet: read the examples)
- Fires on
- 0.69% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 65%
- story 26%
- python 5%
- chat 4%
- worksheet 0%
Makes the model say
ever<|tool_call_end|>Okay me exampleized average will
Makes it not say
" of " like(- love asvalue
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.16. The cake was full of toys! All the kids were so surprised and happy. They played
-
chat 2.0, it turned into colorful paint! Max and his friends were surprised and excited. They continued
-
agent 1.99, question, read, shell, skill, subagent, webfetch
-
agent 1.98, question, read, shell, skill, subagent, webfetch
-
agent 1.97, question, read, shell, skill, subagent, webfetch
-
story 1.96the chair. The baby felt confused and didn't understand why
-
agent 1.95, question, read, shell, skill, subagent, webfetch
-
agent 1.95, question, read, shell, skill, subagent, webfetch
-
agent 1.94, question, read, shell, skill, subagent, webfetch
-
story 1.93and the funny dog all became very small. Tim and Lisa were surprised and didn't know what
-
agent 1.93, question, read, shell, skill, subagent, webfetch
-
agent 1.9, question, read, shell, skill, subagent, webfetch