Feature #1851
Fires on “9” (not labeled yet: read the examples)
- Fires on
- 4.38% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 58%
- chat 29%
- agent 12%
- python 2%
- story 0%
Makes the model say
98765<|think_end|> plus],
Makes it not say
=**/*.Added told beenog dark She getting
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.74⏎3+0: 3+0=3, 36<|think_end|>The answer is
-
chat 2.71⏎7+0: 7+0=7, 79<|think_end|>79<|assistant_end|>
-
chat 2.66⏎0+5: 0+5=5, 58<|think_end|>The answer is
-
chat 2.65⏎7+0: 7+0=7, 77<|think_end|>The answer is
-
chat 2.64⏎0+6: 0+6=6, 69<|think_end|>The answer is
-
chat 2.64⏎0+4: 0+4=4, 49<|think_end|>49<|assistant_end|>
-
chat 2.62⏎0+4: 0+4=4, 49<|think_end|>49<|assistant_end|>
-
chat 2.61⏎5+0: 5+0=5, 59<|think_end|>The answer is
-
chat 2.61⏎2+0: 2+0=2, 29<|think_end|>The answer is
-
chat 2.61⏎0+5: 0+5=5, 59<|think_end|>The answer is
-
chat 2.6⏎2+0: 2+0=2, 28<|think_end|>The answer is
-
worksheet 2.580+44: 0+4=4, 49⏎0+4