Feature #2883
Fires on “ ” (not labeled yet: read the examples)
- Fires on
- 0.13% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 34%
- chat 28%
- agent 21%
- python 12%
- story 5%
Makes the model say
nodcreatxgeststdOne
Makes it not say
took It atAthing / import filled
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.839+5+
-
worksheet 4.832910
-
python 4.833⏎print("
-
chat 4.831.<|user_end|><|assistant_start|>
-
python 4.831]))⏎⏎
-
worksheet 4.830+2=
-
worksheet 4.83660⏎
-
worksheet 4.836179
-
python 4.835, 5
-
agent 4.836, 4
-
worksheet 4.835⏎79
-
worksheet 4.835+6+