Feature #3440
Fires on “ are” (not labeled yet: read the examples)
- Fires on
- 0.22% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 89%
- story 5%
- chat 5%
- python 1%
- worksheet 0%
Makes the model say
to buych bpy project catchOK
Makes it not say
` unexpected branch<|think_start|>D sandEx ball
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 4.28bsearch, write, execute<|system_end|><|system_start|>You are an AI agent running
-
agent 4.26bsearch, write, execute<|system_end|><|system_start|>You are an AI agent running
-
agent 4.26bsearch, write, execute<|system_end|><|system_start|>You are an AI agent running
-
agent 4.25bsearch, write, execute<|system_end|><|system_start|>You are an AI agent running
-
agent 4.25bsearch, write, execute<|system_end|><|system_start|>You are an AI agent running
-
agent 4.25bsearch, write, execute<|system_end|><|system_start|>You are an AI agent running
-
agent 4.23bsearch, write, execute<|system_end|><|system_start|>You are an AI agent running
-
agent 4.21bsearch, write, execute<|system_end|><|system_start|>You are an AI agent running
-
agent 4.21bsearch, write, execute<|system_end|><|system_start|>You are an AI agent running
-
agent 4.2bsearch, write, execute<|system_end|><|system_start|>You are an AI agent running
-
agent 4.2bsearch, write, execute<|system_end|><|system_start|>You are an AI agent running
-
agent 4.2bsearch, write, execute<|system_end|><|system_start|>You are an AI agent running