Feature #3979
Fires on “5” (not labeled yet: read the examples)
- Fires on
- 1.14% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 66%
- story 17%
- chat 12%
- agent 2%
- python 2%
Makes the model say
girl Lily said(" showedmd + moral
Makes it not say
fact One very addsOh Sogged trying
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.6⏎Add 42 and 6.⏎42+06: 2+
-
worksheet 2.6⏎Calculate 78 + 5.⏎78+05: 8+
-
worksheet 2.58⏎Calculate 57 + 5.⏎57+05: 7+
-
worksheet 2.54.<|assistant_end|><|bos|>11 + 85⏎11+85: 1+
-
story 2.53the new skin for Lily. When she was done, she put the new sk
-
worksheet 2.47to know 98 + 8.⏎98+08: 8+
-
worksheet 2.47to know 79 + 3.⏎79+03: 9+
-
worksheet 2.44sum of 41 and 5.⏎41+05: 1+
-
worksheet 2.435.⏎Please add 9 and 6.⏎9+6: 9+
-
worksheet 2.42Please add 52 and 7.⏎52+07: 2+
-
worksheet 2.42⏎Add 53 and 5.⏎53+05: 3+
-
worksheet 2.4to know 69 + 9.⏎69+09: 9+