Feature #2298
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.25% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 85%
- worksheet 6%
- story 4%
- chat 4%
- python 1%
Makes the model say
⏎, with and `. make says
Makes it not say
FixexpressionpreviouspytestTellpa miniRenamed
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.23<|user_end|><|assistant_start|><|think_start|>65+08: 5+8=13, 3
-
chat 3.21<|user_end|><|assistant_start|><|think_start|>18+85: 8+5=13, 3
-
worksheet 3.2now?⏎08+85: 8+5=13, 3
-
chat 3.2<|user_end|><|assistant_start|><|think_start|>65+08: 5+8=13, 3
-
chat 3.18<|user_end|><|assistant_start|><|think_start|>58+95: 8+5=13, 3
-
chat 3.15now?<|user_end|><|assistant_start|><|think_start|>8+5: 8+5=13, 1
-
chat 3.08<|think_start|>538+005: 8+5=13, 3
-
worksheet 3.07= ?⏎98+05: 8+5=13, 3
-
chat 3.06<|user_end|><|assistant_start|><|think_start|>95+88: 5+8=13, 3
-
worksheet 3.0me?⏎08+75: 8+5=13, 3
-
chat 2.99<|user_end|><|assistant_start|><|think_start|>05+58: 5+8=13, 3
-
chat 2.99<|think_start|>005+798: 5+8=13, 3