Feature #13
Fires on “ in” (not labeled yet: read the examples)
- Fires on
- 0.29% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 50%
- chat 23%
- python 11%
- agent 8%
- worksheet 8%
Makes the model say
aym So://ptildtractZ
Makes it not say
answer<|think_end|>," cat<|assistant_end|> belongedvernumber
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.8How can I help you today?<|assistant_end|><|bos|>Once upon a time, in a small town,
-
story 2.79ing colors and playing together.<|bos|>Once upon a time, in a big forest,
-
chat 2.79Can you tell me a story?<|user_end|><|assistant_start|>Once upon a time, in a small town,
-
story 2.799 = 10087<|bos|>Once upon a time, in a big forest,
-
chat 2.79want a story about the sea!<|user_end|><|assistant_start|>Once upon a time, in a big sea,
-
story 2.774 equals 3794.<|bos|>Once upon a time, in an ordinary house
-
story 2.77s<|user_end|><|assistant_start|>You're welcome!<|assistant_end|><|bos|>Once upon a time, in a big forest,
-
story 2.77The answer is 74.<|assistant_end|><|bos|>Once upon a time, in a small town,
-
story 2.77<|bos|>Once upon a time, in a big garden,
-
story 2.76The answer is 14.<|assistant_end|><|bos|>Once upon a time, in a peaceful little town
-
story 2.76The answer is 13.<|assistant_end|><|bos|>Once upon a time, in a bright, sunny
-
story 2.76answer is 498.<|assistant_end|><|bos|>Once upon a time, in a small house,