Feature #1989
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 0.43% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 62%
- chat 27%
- agent 5%
- python 4%
- story 3%
Makes the model say
utils testscase9name InsteadoldStringJack
Makes it not say
to help danced car>>> you " add
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 5.8+4: 4+4=8, 8
-
chat 5.6744+25: 4+5=9, 9
-
worksheet 5.665: 3+5+1=9, 9
-
worksheet 5.45+647: 1+7=8, 8
-
chat 5.4374+295: 4+5=9, 9
-
chat 5.265?<|user_end|><|assistant_start|><|think_start|>4+5: 4+5=9, 9
-
worksheet 5.24834+004: 4+4+1=9, 9
-
worksheet 5.22, 880⏎5+4: 5+4=9, 9
-
worksheet 5.224⏎003+276: 3+6=9, 9
-
chat 5.2: 3+6=9, 9
-
worksheet 5.1695⏎01+67: 1+7=8, 8
-
worksheet 5.14481+927: 1+7+1=9, 9