Feature #1852
Fires on “e” (not labeled yet: read the examples)
- Fires on
- 0.56% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 72%
- story 20%
- chat 5%
- python 2%
- worksheet 1%
Makes the model say
ed ForlpextSorrygest upon
Makes it not say
9 man he turns anymore spotTom i
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 5.26there was a little girl who lived on a gloomy day. She yearned for something
-
agent 5.1bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 5.08bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 5.07bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 5.07bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 5.05bfetch, websearch, write, execute, g
-
agent 5.03bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 5.03bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 5.03bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 5.03bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 5.03bfetch, websearch, write, execute<|system_end|><|system_start|>
-
agent 5.01bfetch, websearch, write, execute<|system_end|><|system_start|>