Feature #1691
Fires on “ob” (not labeled yet: read the examples)
- Fires on
- 0.84% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 48%
- story 33%
- chat 12%
- python 6%
- worksheet 0%
Makes the model say
outsideheadpy suddenly timey length then
Makes it not say
<|assistant_end|> him separoukedoldString⏎ He
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.75glob, grep,
-
agent 2.75glob, grep,
-
agent 2.75glob, grep,
-
agent 2.75glob, grep,
-
agent 2.75glob, grep,
-
agent 2.75glob, grep,
-
agent 2.75glob, grep,
-
agent 2.75glob, grep,
-
agent 2.75glob, grep,
-
agent 2.75glob, grep,
-
story 2.43them a fun surprise.<|bos|>Once there was a soldier named Pete. Pet
-
agent 2.32⏎pizza!⏎```<|assistant_end|><|user_start|>Book a flight to Paris.<|user_end|><|assistant_start|>