Feature #813
Fires on “ their” (not labeled yet: read the examples)
- Fires on
- 0.13% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 56%
- story 32%
- chat 7%
- worksheet 2%
- python 2%
Makes the model say
" hadEvery0 enjoy h takingore
Makes it not say
function Suddenly brother!") watch<|think_end|>SE up
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.12decided to share the fruit with all their jungle friends. Max
-
story 2.01and Willy were very happy, and they shared the treasure with all their friends in the sea
-
story 1.73was so proud of his knowledge. He shared it with all his friends, and he
-
story 1.67he picked the oranges. They shared the oranges with their friends.⏎The
-
story 1.66all day. They were very happy and shared their new treasure with all their friends.<|bos|>Once
-
story 1.6Lily showed her picture to her mom. "Look
-
story 1.55keep it unfolded and shared the wisdom with both of them. They
-
story 1.45can come back and play with me. But he has to ask first before he eats our
-
story 1.45. There was no film inside! Instead, there were yummy treats for all of them. They
-
story 1.44and shared them with their friends. And they
-
story 1.44Everyone in the town loved the festival. It was a special time for all the people.⏎
-
story 1.41river? It is here to protect you, if you don't watch out, the chain