Feature #2329
Fires on “s” (not labeled yet: read the examples)
- Fires on
- 0.78% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 90%
- story 7%
- worksheet 2%
- chat 1%
- python 0%
Makes the model say
def my Mr I* little handtriangle
Makes it not say
+lastular_Amy tests lastv
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 1.7def larger(a, b):⏎⏎⏎script
-
agent 1.65shell<|arg|>command
-
agent 1.65shell<|arg|>command
-
agent 1.65s!⏎<|tool_end|><|assistant_start|>
-
agent 1.65s):⏎ return s
-
python 1.65s):⏎ return s
-
agent 1.65shell<|arg|>command
-
agent 1.65s):⏎2:
-
python 1.65s):⏎ return s
-
agent 1.65shell<|arg|>command
-
agent 1.65s)<|arg|>newString=
-
story 1.65s and barks and