Feature #1158
Fires on “9” (not labeled yet: read the examples)
- Fires on
- 0.66% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 49%
- chat 30%
- agent 17%
- story 4%
- python 0%
Makes the model say
+ worry upononehead rocks<|arg|>odd
Makes it not say
ats they of They tooHowed3
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.0515, 56⏎6+9: 6+9+1=1
-
worksheet 1.9916, 6333⏎7+2: 7+2+1
-
worksheet 1.92, 5101⏎6+9: 6+9+1=1
-
worksheet 1.9115, 5402⏎4+3: 4+3+1
-
worksheet 1.916, 65⏎99+54: 9+4+1
-
worksheet 1.8715, 5586⏎4+5: 4+5+1
-
chat 1.87=16, 6⏎75+89: 5+9+1
-
worksheet 1.8415, 5095⏎4+4: 4+4+1
-
chat 1.847=13, 38⏎4+1: 4+1+1
-
chat 1.831, 1⏎99+69: 9+9+1=1
-
worksheet 1.8316, 6860⏎3+9: 3+9+1
-
chat 1.82=15, 5⏎2+7: 2+7+1=1