Feature #2735
Fires on “ ” (not labeled yet: read the examples)
- Fires on
- 0.11% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 46%
- chat 33%
- story 11%
- python 10%
- agent 1%
Makes the model say
returnspy her intoMy): bothOnce
Makes it not say
ue function ==Mom`, Mom noise helpers
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.03⏎8+7: 8+7=15, 1592
-
worksheet 2.964⏎7+8: 7+8=15, 1581
-
chat 2.95+8: 8+8=16, 167<|think_end|>
-
worksheet 2.958+8: 8+8+1=17, 1791
-
worksheet 2.923⏎9+7: 9+7=16, 1632
-
chat 2.877+8: 7+8+1=16, 1656
-
chat 2.878+7: 8+7+1=16, 1613
-
chat 2.878+8: 8+8+1=17, 1706
-
worksheet 2.869⏎7+9: 7+9=16, 1671
-
worksheet 2.857⏎9+6: 9+6=15, 1595
-
worksheet 2.850⏎9+8: 9+8=17, 1758
-
chat 2.837+8: 7+8+1=16, 1656