Feature #522
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 0.60% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 50%
- agent 30%
- chat 9%
- story 6%
- python 5%
Makes the model say
return returns or ? I = had thanked
Makes it not say
mix too're first betterWhile sum learn
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 4.03121949.⏎6 + 810 = 816
-
worksheet 3.95242+04573: 2+3=5, 5
-
chat 3.93<|assistant_start|><|think_start|>343+212: 3+2=5, 5
-
worksheet 3.820002+4193: 2+3=5, 5
-
chat 3.81?<|user_end|><|assistant_start|><|think_start|>22+03: 2+3=5, 5
-
chat 3.79<|assistant_start|><|think_start|>743+422: 3+2=5, 5
-
chat 3.78?<|user_end|><|assistant_start|><|think_start|>93+02: 3+2=5, 5
-
chat 3.76<|assistant_start|><|think_start|>852+443: 2+3=5, 5
-
worksheet 3.728⏎723+092: 3+2=5, 5
-
chat 3.72<|assistant_start|><|think_start|>393+052: 3+2=5, 5
-
worksheet 3.72122+98333: 2+3=5, 5
-
worksheet 3.71⏎003+482: 3+2=5, 5