Feature #1892
Fires on “<|system_start|>” (not labeled yet: read the examples)
- Fires on
- 0.28% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 86%
- worksheet 7%
- chat 6%
- story 2%
- python 0%
Makes the model say
you best now barkeded most yours
Makes it not say
function ():`# defines laptop average
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.474.⏎If you add 9 and 840, you get 84
-
agent 2.37, websearch, write, execute<|system_end|><|system_start|>You are an AI
-
agent 2.35, websearch, write, execute<|system_end|><|system_start|>You are an AI
-
agent 2.35, websearch, write, execute<|system_end|><|system_start|>You are an AI
-
worksheet 2.348.⏎If you add 9 and 876, you get 88
-
agent 2.33, websearch, write, execute<|system_end|><|system_start|>You are an AI
-
agent 2.33, websearch, write, execute<|system_end|><|system_start|>You are an AI
-
worksheet 2.334.⏎If you add 1 and 341, you get 34
-
agent 2.32, websearch, write, execute<|system_end|><|system_start|>You are an AI
-
agent 2.32, websearch, write, execute<|system_end|><|system_start|>You are an AI
-
agent 2.32, websearch, write, execute<|system_end|><|system_start|>You are an AI
-
agent 2.31, websearch, write, execute<|system_end|><|system_start|>You are an AI