Feature #2524
Fires on “ ex” (not labeled yet: read the examples)
- Fires on
- 0.52% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 77%
- story 17%
- chat 3%
- python 3%
- worksheet 0%
Makes the model say
erhhellomangisttraor
Makes it not say
py]) she it")+ her```
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 5.77wanted. He learned his lesson and now always makes sure he gives himself extra time in case
-
agent 5.72webfetch, websearch, write, execute<|system_end|>
-
agent 5.69webfetch, websearch, write, execute<|system_end|>
-
agent 5.67webfetch, websearch, write, execute<|system_end|>
-
agent 5.63webfetch, websearch, write, execute<|system_end|>
-
agent 5.61webfetch, websearch, write, execute<|system_end|>
-
agent 5.59webfetch, websearch, write, execute<|system_end|>
-
agent 5.59webfetch, websearch, write, execute<|system_end|>
-
agent 5.59webfetch, websearch, write, execute<|system_end|>
-
agent 5.58webfetch, websearch, write, execute<|system_end|>
-
agent 5.58webfetch, websearch, write, execute,
-
agent 5.56webfetch, websearch, write, execute<|system_end|>