Feature #3672
Fires on “ an” (not labeled yet: read the examples)
- Fires on
- 0.37% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 77%
- story 11%
- chat 11%
- python 1%
- worksheet 0%
Makes the model say
printa sister B good seek liked dog
Makes it not say
edit<|tool_start|> wish⏎ iceSetreturn told
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.69search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 3.68search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 3.68search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 3.68search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 3.67search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 3.66search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 3.66search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 3.66search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 3.66search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 3.66search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 3.66search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 3.65search, write, execute<|system_end|><|system_start|>You are an AI agent running in