Feature #1809
Fires on “<|bos|>” (not labeled yet: read the examples)
- Fires on
- 1.03% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 42%
- chat 28%
- python 13%
- agent 10%
- worksheet 7%
Makes the model say
One wanted Lucy liked ground Tom Tim +
Makes it not say
too himselfR whyace so$But
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.61<|bos|>Once upon a time
-
story 2.61<|bos|>Once upon a time
-
python 2.61<|bos|>Bug: `max
-
story 2.61<|bos|>Once upon a time
-
chat 2.61<|bos|><|user_start|>What are you
-
chat 2.61<|bos|><|system_start|>Answer in one
-
agent 2.61<|bos|><|system_start|>tools: edit
-
story 2.61<|bos|>Once upon a time
-
chat 2.61<|bos|><|system_start|>tools: calculator
-
story 2.61<|bos|>Once upon a time
-
story 2.61<|bos|>Once upon a time
-
chat 2.61<|bos|><|user_start|>Calculate 5