Feature #490
Fires on “ can” (not labeled yet: read the examples)
- Fires on
- 1.38% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 35%
- chat 25%
- story 19%
- worksheet 12%
- python 8%
Makes the model say
yourdy prelude accRenamed yourself need[-
Makes it not say
and0 fromco45:` went
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.43can you help me?<|user_end|>
-
agent 3.43can you fix it?<|user_end|>
-
agent 3.43can you fix it?<|user_end|>
-
story 3.43can you help me shake the
-
story 3.35help them if they don't know
-
story 3.33play in the snow with
-
story 2.82can be free and play
-
agent 2.82can list, read,
-
story 2.82can do special things."
-
chat 2.82can only tell short stories
-
chat 2.82can only tell short stories
-
chat 2.82can't help with that