Feature #630
Fires on “'m” (not labeled yet: read the examples)
- Fires on
- 0.59% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 40%
- story 36%
- chat 24%
- worksheet 0%
- python 0%
Makes the model say
a sorrymain so red result welcome is
Makes it not say
]`. iceOG[- breakidentible
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 4.46and says, "That's very kind of you, Anna. I'm sure the rabbit will
-
story 4.38and sad. He says, "I'm sorry, Mom,
-
story 4.33mommy hugged her and said, "You're welcome, Mia. I'm proud of you for
-
chat 4.15Can you book a table for two tonight?<|user_end|><|assistant_start|>I'm sorry, I can
-
story 4.11tie. He said, "I'm sorry I ran away
-
chat 4.1said, "I am not crazy, I
-
story 4.05thank you!"⏎His Mom hugged him back and said, "I'm glad you like it
-
story 4.04your egg, and I am so sorry." Bob
-
story 4.04started to talk! The tree said, "I am the memory
-
story 4.03Mom!"⏎Mom says, "You're welcome, Ben. I'm glad you found it
-
story 4.0clown smiled and said, "You're welcome, my friend. I'm glad I could make
-
story 3.98she said, "Thank you, Fred and Fiona! I am a magic bird,