Feature #839
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 1.55% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 26%
- chat 24%
- python 21%
- story 21%
- worksheet 8%
Makes the model say
= testsOncelowerutils Itnder there
Makes it not say
ue listen [ and while<|arg|> like([
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.928⏎5+1: 5+1+1=7, 7
-
chat 2.8, 8⏎5+1: 5+1+1=7, 7
-
chat 2.5106⏎5+1: 5+1+1=7, 7
-
chat 2.510⏎05+71: 5+1+1=7, 7
-
chat 2.5, 8⏎5+1: 5+1+1=7, 7
-
chat 2.46.<|assistant_end|><|user_start|>And if I add 5
-
chat 2.43?<|user_end|><|assistant_start|><|think_start|>71+06: 1+6=7, 7
-
worksheet 2.410⏎61+06: 1+6=7, 7
-
agent 2.38.<|assistant_end|><|bos|><|system_start|>You
-
story 2.38. He sat down at
-
story 2.38. She sees the o
-
chat 2.38. I only know how