Feature #1273
Fires on “ Mia” (not labeled yet: read the examples)
- Fires on
- 0.31% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 45%
- worksheet 24%
- story 15%
- agent 12%
- python 4%
Makes the model say
plus adds himdemovalue woke countsIt
Makes it not say
`⏎ explore manyFind lot climbeditem
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.21ave a nice day.<|assistant_end|><|bos|>Tim has 86 shells. Tim gets 72
-
chat 2.9800.<|assistant_end|><|user_start|>Sue has 91 coins. Sue gets 54
-
worksheet 2.8618591⏎Tom has 13 cookies. Tom gets 33
-
chat 2.79use the calculator.<|system_end|><|user_start|>Mia has 917 stickers. Mia gets 0 more
-
chat 2.76tests pass now.<|assistant_end|><|bos|><|user_start|>Lily has 71 coins. Lily gets 8 more
-
chat 2.75<|think_end|>549<|assistant_end|><|bos|><|user_start|>Tim has 19 cookies. Tim gets 76
-
worksheet 2.7579424.⏎Mia has 8 books. Mia gets 85
-
chat 2.7is 28.<|assistant_end|><|user_start|>Lily has 47 books. Lily gets 45
-
chat 2.7Do not use the calculator.<|system_end|><|user_start|>Mia has 8 stickers. Mia gets 2 more
-
chat 2.676618.<|bos|><|user_start|>Lucy has 18 coins. Lucy gets 79
-
worksheet 2.665100.⏎Mia has 823 apples. Mia gets 60
-
chat 2.64answer is 81.<|assistant_end|><|user_start|>Tim has 2 balls. Tim gets 8 more