Feature #1732
Fires on “<|bos|>” (not labeled yet: read the examples)
- Fires on
- 1.11% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 30%
- story 25%
- agent 24%
- worksheet 11%
- python 10%
Makes the model say
9 tall returns there= ==py"))
Makes it not say
<|assistant_end|> later uponThank<|arg|> whenood From
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 22.38they had a fun day playing with the treasure in the park.<|assistant_end|><|bos|>Once upon a time
-
chat 22.2that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|user_start|>Who made you
-
story 22.14that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|>Once upon a time
-
story 22.09make everyone happy, even when things don't go as planned.<|bos|>Once upon a time
-
agent 22.07tiny model: I can only tell short stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit
-
agent 22.04that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit
-
agent 22.02on, they had many fun adventures.<|assistant_end|><|bos|><|system_start|>tools: edit
-
chat 22.02about that. I can only tell short stories and add numbers.<|assistant_end|><|bos|><|user_start|>Can you write
-
chat 22.02that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>Start every answer
-
worksheet 22.01that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|>Lily has 3
-
agent 22.01that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit
-
agent 22.01that. I only know how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit