Feature #1572
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.64% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 76%
- chat 21%
- story 1%
- agent 1%
- python 1%
Makes the model say
7356es04 greeting
Makes it not say
`https Come("_[::-:Fix
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 5.599.⏎If you add 66972 and 19, you get
-
worksheet 5.5.<|assistant_end|><|bos|>If you add 76957 and 12, you get
-
worksheet 5.4926.⏎If you add 9238 and 211, you
-
worksheet 5.26960.⏎Add 12742 and 188.⏎
-
worksheet 5.245.⏎If you add 11553 and 3870,
-
worksheet 5.2497.⏎If you add 8129 and 52446
-
worksheet 5.21⏎If you add 47536 and 8, you get
-
worksheet 5.166.⏎If you add 44764 and 2, you get
-
worksheet 5.084.⏎Tim had 3946 flowers and found 31929
-
worksheet 5.0699.⏎If you add 9674 and 350, you
-
worksheet 5.022.<|assistant_end|><|bos|>If you add 8569 and 7673,
-
worksheet 5.0269.⏎Lucy had 298 books and found 5 more. Now Lucy