Feature #296
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.32% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 50%
- chat 29%
- story 11%
- agent 9%
- python 2%
Makes the model say
ay while wip other more years prelude leave
Makes it not say
out.ce for +helloThe three
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.17<|think_start|>9+9: 9+9=18, 18<|think_end|>The answer
-
chat 3.14<|think_start|>8+9: 8+9=17, 17<|think_end|>The answer
-
chat 2.99+9: 4+9+1=14, 145<|think_end|>The
-
chat 2.97<|think_start|>8+9: 8+9=17, 17<|think_end|>The answer
-
worksheet 2.91+8: 9+8+1=18, 18440
-
chat 2.85<|think_start|>8+5: 8+5=13, 13<|think_end|>13
-
worksheet 2.84+8: 9+8+1=18, 18291
-
chat 2.83⏎8+9: 8+9=17, 175<|think_end|>1
-
chat 2.83<|think_start|>9+8: 9+8=17, 17<|think_end|>The answer
-
worksheet 2.82⏎9+8: 9+8=17, 17580
-
worksheet 2.81+5: 9+5+1=15, 15465
-
chat 2.79+9: 4+9+1=14, 1406<|think_end|>