Feature #2303
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.43% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 51%
- chat 40%
- story 6%
- python 2%
- agent 1%
Makes the model say
help s all, funny donain one
Makes it not say
path<|assistant_end|>_*. ushpattern defines closer
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.42⏎5+5: 5+5=10, 10
-
chat 3.42, 8⏎6+5: 6+5=11, 11
-
worksheet 3.235+556: 5+6=11, 1⏎
-
chat 3.2148⏎5+7: 5+7=12, 12
-
chat 3.18145+006: 5+6=11, 1⏎
-
worksheet 3.17?⏎66+25: 6+5=11, 1⏎
-
chat 3.04, 7⏎6+5: 6+5=11, 11
-
worksheet 2.996+00005: 6+5=11, 1⏎
-
chat 2.94.<|user_end|><|assistant_start|><|think_start|>7+3: 7+3=10, 10
-
chat 2.945: 5+5=10, 10
-
worksheet 2.896.⏎5+6: 5+6=11, 11
-
chat 2.86?<|user_end|><|assistant_start|><|think_start|>2+8: 2+8=10, 10