Feature #3214
Fires on “9” (not labeled yet: read the examples)
- Fires on
- 0.65% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 87%
- worksheet 7%
- story 2%
- chat 2%
- python 2%
Makes the model say
<|tool_end|><|bos|> listenedpspy<|assistant_end|> branch test
Makes it not say
avereturnwWhy why what# mouth
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.899835=⏎
-
worksheet 2.769804.⏎
-
worksheet 2.6494⏎00
-
worksheet 2.64999⏎0
-
worksheet 2.649+448
-
worksheet 2.649+4=1
-
worksheet 2.649⏎53+
-
worksheet 2.649 plus 55
-
python 2.649⏎print("all
-
worksheet 2.6497⏎So
-
agent 2.649: def echo(
-
chat 2.6494<|think_end|>59