Feature #2810
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.47% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 71%
- chat 20%
- python 4%
- story 3%
- agent 2%
Makes the model say
Fixjson searched9txtnder=N
Makes it not say
mini day<|assistant_end|> import Jif typo'm
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.53+ 70 = 88.⏎2+1555=⏎0
-
worksheet 2.44+ 23 = 71.⏎3+4989=⏎0
-
worksheet 2.08= 1560.⏎5577+5224=⏎
-
worksheet 2.06= 46566.⏎34+25375=⏎
-
worksheet 2.0491 = 1174.⏎0+35158=⏎
-
worksheet 1.986 = 317.⏎5859+3928=⏎
-
worksheet 1.9878 = 81.⏎78+730=⏎07
-
worksheet 1.94+ 2 = 6197.⏎4+75=⏎04
-
chat 1.89<|assistant_start|>The answer is 57.<|assistant_end|><|user_start|>0+121=<|user_end|><|assistant_start|>The
-
chat 1.88<|user_end|><|assistant_start|>The answer is 5.<|assistant_end|><|user_start|>8+741=<|user_end|><|assistant_start|>The
-
worksheet 1.83+ 1 = 4.⏎4206+81658=
-
worksheet 1.8248914.⏎4963+20=⏎49