Feature #3171
Fires on “ill” (not labeled yet: read the examples)
- Fires on
- 1.08% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 66%
- agent 18%
- chat 14%
- python 1%
- worksheet 0%
Makes the model say
friendspy'sa([ loved toy proud
Makes it not say
<|assistant_start|>vowels But<|tool_end|>",`,<|tool_call_end|> trained
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 5.16, little rabbit named Ralph. He lived
-
story 4.89.<|assistant_end|><|bos|>Once upon a time, there were two friends called Drew and Jack. Every
-
story 4.76Millie was very happy
-
story 4.7was a little girl named June. June was 3
-
story 4.68when a big wind came. The wind was so strong it blew Mittens up into a
-
story 4.61.⏎One day, Blue had an idea. He said to Red, "Let's
-
story 4.57things to yourself.<|bos|>Once there was a pretty little girl called Millie. One day
-
story 4.55to tell stories to his friends in the woods. One day, Foxy found a
-
story 4.5They were patient and kind and she loved playing with them. But Joy didn't think
-
story 4.49, there was a little car named Speedy. Speedy loved to go
-
story 4.45They smiled and hugged the red train. They decided to name it Ruby. They played
-
story 4.459.<|bos|>Once upon a time there was a little girl called Millie. She was