Feature #560
Fires on “4” (not labeled yet: read the examples)
- Fires on
- 0.19% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 62%
- chat 14%
- agent 12%
- python 7%
- story 4%
Makes the model say
'.iveOK Python`. [ient
Makes it not say
quiet So't sure walkedWhystwhisper
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.784+82:
-
chat 4.7844+25
-
worksheet 4.7845⏎So
-
agent 4.784: print(mean
-
worksheet 4.7847+03
-
chat 4.784.<|assistant_end|><|user_start|>what
-
chat 4.78495<|assistant_end|><|bos|>
-
chat 4.784=11,
-
worksheet 4.7843.⏎6
-
chat 4.784<|tool_end|><|assistant_start|>The answer
-
agent 4.784: acc =
-
python 4.784, 0]