Feature #1021
Fires on “ an” (not labeled yet: read the examples)
- Fires on
- 0.33% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 84%
- story 7%
- python 6%
- chat 3%
- worksheet 1%
Makes the model say
fe away share sharinggerero sharedlength
Makes it not say
# Sheice arecase sheprintecho
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.02surprised because it hadn't taken long at
-
story 1.92sighed. She understood that she would have to wait until she was grown-up,
-
agent 1.89a, b):⏎6: return (a + b) /
-
agent 1.88, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 1.88, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 1.88, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 1.88, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 1.88, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 1.88, execute<|system_end|><|system_start|>You are an AI agent running in
-
agent 1.88, b):⏎6: return (a + b) /
-
agent 1.86⏎8: return (a + b) /
-
agent 1.83write, execute<|system_end|><|system_start|>You are an AI agent running in