Feature #3726
Fires on “3” (not labeled yet: read the examples)
- Fires on
- 0.30% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 35%
- worksheet 25%
- story 23%
- chat 12%
- python 5%
Makes the model say
at,s leaves windelil pet
Makes it not say
<|arg|>vingved edit start boy allarith
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 2.45333333333333333⏎assert mean_
-
worksheet 1.933344 =
-
python 1.933)⏎-3
-
chat 1.93325<|assistant_end|><|bos|>
-
worksheet 1.93312:
-
worksheet 1.933 + 39
-
worksheet 1.9338 + 9
-
worksheet 1.933+181
-
worksheet 1.93347:
-
chat 1.9335.<|user_end|><|assistant_start|>
-
worksheet 1.933515⏎
-
agent 1.933, 7,