Feature #426
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 1.10% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 63%
- agent 22%
- chat 11%
- story 3%
- python 1%
Makes the model say
new function double bitThank!path forgive
Makes it not say
It Heren Igage hid They
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.1806: 4+6+1=11, 13⏎35
-
worksheet 1.987: 8+7+1=16, 60⏎0+2
-
chat 1.983+38: 3+8=11, 1⏎0+3
-
worksheet 1.987995: 6+5=11, 1⏎000
-
worksheet 1.97: 5+7+1=13, 13250
-
worksheet 1.9747: 3+7+1=11, 11⏎95
-
story 1.97danced. She was very enthusiastic when she danced
-
worksheet 1.968: 9+8+1=18, 864⏎61
-
worksheet 1.9469: 1+9+1=11, 10⏎05
-
agent 1.92I help with your code?<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep,
-
worksheet 1.923: 7+3+1=11, 119⏎6
-
agent 1.91!<|assistant_end|><|bos|><|system_start|>tools: edit, glob, grep,