Feature #370
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.33% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 65%
- chat 25%
- agent 6%
- story 4%
- python 1%
Makes the model say
= through // intoighed carefully *echo
Makes it not say
boy pBen so Ben lists talking fish
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 5.521⏎00+97: 0+7+1=8,
-
chat 5.4, 02⏎7+2: 7+2+1=10,
-
worksheet 5.32, 0⏎0+3: 0+3+1=4,
-
worksheet 5.2939⏎47+72: 7+2+1=10,
-
worksheet 5.2306⏎6+2: 6+2+1=9,
-
worksheet 5.18, 50⏎7+0: 7+0+1=8,
-
chat 5.176, 6⏎1+3: 1+3+1=5,
-
chat 5.151, 1⏎7+0: 7+0+1=8,
-
chat 5.14, 03⏎0+7: 0+7+1=8,
-
chat 5.126, 6⏎0+7: 0+7+1=8,
-
chat 5.124⏎37+01: 7+1+1=9,
-
worksheet 5.13⏎01+63: 1+3+1=5,