Feature #3135
Fires on “ gets” (not labeled yet: read the examples)
- Fires on
- 0.17% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 42%
- worksheet 27%
- python 11%
- story 10%
- agent 9%
Makes the model say
s him+ next ` hasion
Makes it not say
Read import<|user_start|>{`.Answer couldTrue
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.37not use the calculator.<|system_end|><|user_start|>Mia has 8 stickers. Mia gets 2 more stickers
-
chat 4.29not use the calculator.<|system_end|><|user_start|>Mia has 8 stickers. Mia gets 2 more stickers
-
chat 3.98to plus".<|assistant_end|><|bos|><|user_start|>Anna has 87 stars. Anna gets 872
-
chat 3.9756.<|assistant_end|><|bos|><|user_start|>Tim has 74 cookies. Tim gets 0 more cookies
-
chat 3.95107<|assistant_end|><|bos|><|user_start|>Sam has 19 toys. Sam gets 38 more
-
chat 3.95<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Mia has 2 coins. Mia gets 6 more coins
-
chat 3.94the calculator.<|system_end|><|user_start|>Mia has 917 stickers. Mia gets 0 more stickers
-
chat 3.93767<|assistant_end|><|bos|><|user_start|>Sam has 63 coins. Sam gets 98 more
-
chat 3.91121.<|assistant_end|><|user_start|>Lucy has 7 books. Lucy gets 2 more books
-
worksheet 3.9a nice day.<|assistant_end|><|bos|>Tim has 86 shells. Tim gets 725
-
chat 3.88549<|assistant_end|><|bos|><|user_start|>Tim has 19 cookies. Tim gets 76 more
-
chat 3.88<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Mia has 0 coins. Mia gets 4 more coins