Feature #1538
Fires on “ a” (not labeled yet: read the examples)
- Fires on
- 0.84% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 65%
- story 18%
- chat 9%
- python 5%
- worksheet 3%
Makes the model say
nod Pythonus squareient[::: learn
Makes it not say
was from they5<|assistant_end|>4ater other
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.34a cake.<|user_end|><|assistant_start|>
-
story 3.34a curious little boy named
-
story 3.34a good helper!"
-
agent 3.34a coding agent harness.
-
chat 3.34a faraway place
-
story 3.34a big smile and loved
-
story 3.34a little cat, said
-
story 3.34a little scary at first
-
agent 3.34a coding agent harness.
-
chat 3.34a fair princess. She
-
story 3.34a rare, yummy treat
-
agent 3.34a function mean(a