Feature #1730
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 1.21% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 50%
- story 24%
- chat 22%
- python 2%
- agent 2%
Makes the model say
<|bos|> he unexpected, that fullone years
Makes it not say
typo to pulled r water( pullCreated
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 4.36288+053: 8+3=11, 1⏎
-
worksheet 4.1812+8858: 2+8=10, 0⏎
-
chat 4.137⏎02+59: 2+9=11, 17
-
worksheet 4.12+548: 8+8=16, 65
-
chat 4.12<|assistant_start|><|think_start|>19+62: 9+2=11, 1⏎
-
chat 4.09, 8⏎6+5: 6+5=11, 11
-
worksheet 4.06?⏎66+25: 6+5=11, 1⏎
-
worksheet 4.032+19178: 2+8=10, 0⏎
-
chat 4.02<|assistant_start|><|think_start|>12+49: 2+9=11, 1⏎
-
worksheet 4.023⏎05+56: 5+6=11, 19
-
worksheet 4.0168+1608: 8+8=16, 6⏎
-
chat 4.01+639: 7+9=16, 6⏎