Feature #2402
Fires on “<|bos|>” (not labeled yet: read the examples)
- Fires on
- 3.26% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 46%
- chat 19%
- python 18%
- worksheet 9%
- agent 8%
Makes the model say
=<|user_end|> wanted( likeddell morning
Makes it not say
This found them False result theyin home
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 6.8737 = 91540<|bos|>Tom and Mia were best friends. They
-
story 6.84<|bos|>Once upon a time
-
story 6.84<|bos|>Once upon a time
-
python 6.84<|bos|>Bug: `max
-
story 6.84<|bos|>Once upon a time
-
chat 6.84<|bos|><|user_start|>What are you
-
chat 6.84<|bos|><|system_start|>Answer in one
-
agent 6.84<|bos|><|system_start|>tools: edit
-
story 6.84<|bos|>Once upon a time
-
chat 6.84<|bos|><|system_start|>tools: calculator
-
story 6.84<|bos|>Once upon a time
-
story 6.84<|bos|>Once upon a time