Feature #1905
Fires on “ How” (not labeled yet: read the examples)
- Fires on
- 0.12% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 35%
- worksheet 32%
- agent 16%
- story 14%
- python 3%
Makes the model say
readmeanatedpowerAmyaavehing
Makes it not say
# hill glue down But ear sun The
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.9887 stars. Anna gets 872 more stars. How many stars does Anna
-
chat 4.95has 42 balls. Sue gets 18 more balls. How many balls does Sue
-
chat 4.9has 63 coins. Sam gets 98 more coins. How many coins does Sam
-
chat 4.85has 41 toys. Anna gets 37 more toys. How many toys does Anna
-
chat 4.8455 coins. Sue gets 303 more coins. How many coins does Sue
-
chat 4.83has 41 stars. Tim gets 88 more stars. How many stars does Tim
-
worksheet 4.746 shells. Tim gets 72522 more shells. How many shells does Tim
-
worksheet 4.737274 apples. Tom gets 39 more apples. How many apples does Tom
-
chat 4.7272 stickers. Sue gets 651 more stickers. How many stickers does Sue
-
chat 4.71<|user_start|>Tim has 2 balls. Tim gets 8 more balls. How many balls does Tim
-
chat 4.67has 63 coins. Sam gets 98 more coins. How many coins does Sam
-
chat 4.65818 flowers. Sam gets 71 more flowers. How many flowers does Sam