Feature #3494
Fires on “0” (not labeled yet: read the examples)
- Fires on
- 1.57% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 60%
- chat 29%
- python 8%
- agent 2%
- story 1%
Makes the model say
: tests):name Try= Max waved
Makes it not say
odd were fisicm felt feel
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 4.087⏎000+600: 0+
-
worksheet 4.049+4487: 9
-
worksheet 3.7713.⏎42+5=⏎42+05: 2+
-
worksheet 3.725⏎794+943: 4+
-
chat 3.67<|think_start|>03+24: 3+
-
worksheet 3.64together.<|bos|>Compute 49+0⏎49+00: 9+
-
worksheet 3.590+8=8,
-
worksheet 3.585.⏎Compute 39+3⏎39+03: 9+
-
worksheet 3.5696.⏎46+0=⏎46+00: 6+
-
worksheet 3.56⏎481+927: 1+
-
chat 3.565+272: 5+
-
chat 3.55<|assistant_end|><|user_start|>Plus 6?<|user_end|><|assistant_start|><|think_start|>71+06: 1+