Feature #2700
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.40% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 32%
- chat 23%
- agent 19%
- python 19%
- story 6%
Makes the model say
1commandcase0 Falsebed]`. functions
Makes it not say
and= from his of_!math
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 4.4⏎5+5: 5+5=10, 1
-
chat 4.345: 5+5=10, 1
-
chat 4.11<|user_end|><|assistant_start|><|think_start|>05+45: 5+5=10, 0
-
worksheet 3.979286⏎5+5: 5+5=10, 1
-
worksheet 3.88801⏎5+5: 5+5=10, 1
-
chat 3.844⏎05+45: 5+5=10, 0
-
worksheet 3.84⏎035+515: 5+5=10, 0
-
worksheet 3.83965+4685: 5+5=10, 0
-
worksheet 3.7865+00585: 5+5=10, 0
-
worksheet 3.77755+5295: 5+5=10, 0
-
chat 3.73<|think_start|>055+715: 5+5=10, 0
-
worksheet 3.715, 5⏎5+5: 5+5=10, 1