Feature #131
Fires on “ a” (not labeled yet: read the examples)
- Fires on
- 0.49% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 54%
- agent 26%
- chat 10%
- story 9%
- worksheet 2%
Makes the model say
a dog cat realT w k-
Makes it not say
([thWhat whatreplaceAll first H test
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.01a cake.<|user_end|><|assistant_start|>
-
story 4.01a curious little boy named
-
story 4.01a good helper!"
-
agent 4.01a coding agent harness.
-
chat 4.01a faraway place
-
story 4.01a big smile and loved
-
story 4.01a little cat, said
-
story 4.01a little scary at first
-
agent 4.01a coding agent harness.
-
chat 4.01a fair princess. She
-
story 4.01a rare, yummy treat
-
agent 4.01a function mean(a