Feature #2984
Fires on “ an” (not labeled yet: read the examples)
- Fires on
- 0.94% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 58%
- story 33%
- chat 6%
- worksheet 3%
- python 1%
Makes the model say
brave enough and even ataverage should share
Makes it not say
=:case + noises<|assistant_end|>", "!"
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 5.2search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 5.18search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 5.14search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 5.13search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 5.13search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 5.13search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 5.11search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 5.09search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 5.09search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 5.09search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 5.09search, write, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 5.08search, write, execute<|system_end|><|system_start|>You are an AI agent running in