Feature #872
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 0.61% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 42%
- chat 22%
- agent 17%
- worksheet 14%
- python 4%
Makes the model say
's. =ant?"negateent beak
Makes it not say
enough$This toys surprise smiledaidschema
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.7817⏎1+2: 1+2+1=4, 4
-
story 2.76Tim asked his mom, "Can you help me get the cone?" His mom tried to reach
-
worksheet 2.65382+071: 2+1+1=4, 4
-
worksheet 2.64281+412: 1+2+1=4, 4
-
chat 2.61<|assistant_start|><|think_start|>992+092: 2+2=4, 4
-
chat 2.5917⏎1+2: 1+2+1=4, 4
-
chat 2.5972⏎2+1: 2+1+1=4, 4
-
worksheet 2.577⏎81+62: 1+2+1=4, 4
-
chat 2.54<|assistant_start|><|think_start|>002+762: 2+2=4, 4
-
chat 2.52<|assistant_start|><|think_start|>452+142: 2+2=4, 4
-
worksheet 2.52, 1⏎1+2: 1+2+1=4, 4
-
worksheet 2.53⏎412+022: 2+2=4, 4