Feature #2465
Fires on “7” (not labeled yet: read the examples)
- Fires on
- 1.10% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 84%
- chat 7%
- story 4%
- agent 4%
- python 2%
Makes the model say
sad7<|arg|> insteadde goodbye careful lost
Makes it not say
! things Inside them out on music
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 1.75+8=16, 6⏎706+160: 6
-
worksheet 1.71+4=5, 59⏎330+101: 0
-
worksheet 1.651=16, 65⏎815+138: 5
-
worksheet 1.58=14, 4⏎2254+7194: 4
-
worksheet 1.58+5=13, 3⏎093+174: 3
-
worksheet 1.579=17, 7⏎982+016: 2+
-
worksheet 1.568=13, 33⏎791+271: 1
-
worksheet 1.56+1=8, 84⏎002+179: 2
-
worksheet 1.55+9=9, 99⏎093+679: 3
-
worksheet 1.541=11, 10⏎050+976: 0
-
worksheet 1.531=12, 23⏎000+181: 0
-
worksheet 1.52+5=13, 3⏎625+001: 5