Feature #2706
Fires on “ with” (not labeled yet: read the examples)
- Fires on
- 1.03% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 47%
- agent 23%
- chat 12%
- python 12%
- worksheet 6%
Makes the model say
<|bos|> the sorry= +( %[-
Makes it not say
first working strings another moralstrings"atter
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 5.32by step or with a calculator.<|assistant_end|>
-
agent 5.18with a function hello that
-
story 5.18with a long string.
-
story 5.18with me?"⏎Tommy
-
story 5.18with an arrow on it
-
chat 5.18with "Sure!".<|system_end|>
-
story 5.18with many birds on it
-
chat 5.18with the number only.
-
story 5.18with the sticks. He
-
chat 5.18with the number only.
-
chat 5.18with the lemon. It
-
agent 5.18with `final_letter