Feature #1708
Fires on “9” (not labeled yet: read the examples)
- Fires on
- 0.92% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 38%
- story 34%
- worksheet 13%
- chat 10%
- python 4%
Makes the model say
proud<|arg|>ight surenumberoldString<|system_end|> matter
Makes it not say
veryir named HerassertHi allc
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.69999⏎0+
-
worksheet 2.6999835=
-
worksheet 2.6999: 0+
-
worksheet 2.6994⏎00
-
worksheet 2.699+448
-
worksheet 2.699+4=1
-
worksheet 2.699⏎53+
-
worksheet 2.699804.
-
worksheet 2.699 plus 55
-
python 2.699⏎print("all
-
worksheet 2.6997⏎So
-
agent 2.699: def echo(