Feature #1923
Fires on “ ” (not labeled yet: read the examples)
- Fires on
- 0.62% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 64%
- chat 35%
- story 0%
- agent 0%
- python 0%
Makes the model say
insidele bestgreet( lot([ later
Makes it not say
And wanted broke getspositive$ *<|tool_call_start|>
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.72+1=15, 56⏎7+8: 7+8+
-
chat 3.68+1=19, 91⏎9+6: 9+6+
-
chat 3.67+1=10, 06⏎4+9: 4+9+
-
chat 3.66+1=15, 56⏎7+8: 7+8+
-
chat 3.66+1=19, 91⏎9+6: 9+6+
-
worksheet 3.65+1=19, 95⏎1+3: 1+3+
-
worksheet 3.64+1=13, 34⏎7+7: 7+7+
-
chat 3.64+1=17, 75⏎6+6: 6+6+
-
worksheet 3.621=14, 465⏎9+5: 9+5+
-
chat 3.612+9=11, 1⏎4+7: 4+7+
-
worksheet 3.61=17, 7299⏎0+6: 0+6+
-
worksheet 3.6+9=18, 83⏎4+6: 4+6+