Feature #3535
Fires on “One” (not labeled yet: read the examples)
- Fires on
- 0.08% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 71%
- chat 24%
- agent 3%
- worksheet 1%
- python 1%
Makes the model say
uponOnce day# For pythonually plus
Makes it not say
each heone⏎rednumberdouble
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 5.8682347 apples.<|bos|>One day, a little
-
story 5.76+ 73 = 82.<|bos|>One day, a little
-
story 5.5955499 = 57077.<|bos|>One day, a fast
-
story 5.5529839 = 74763.<|bos|>One day, a little
-
story 5.51843.<|assistant_end|><|bos|>One day, a boy
-
story 5.47Now Lucy has 11626 toys.<|bos|>One day, a smart
-
story 5.462 + 475 = 477.<|bos|>One day, a little
-
story 5.453 + 769 = 772.<|bos|>One day, a little
-
story 5.42⏎So 4 + 89 = 93.<|bos|>One day, a big
-
story 5.374 + 907 = 1021.<|bos|>One day, a little
-
story 5.3470 + 2 = 82672.<|bos|>One day, Joey
-
story 5.3+ 6385 = 8745.<|bos|>One day, a little