Feature #119
Fires on “7” (not labeled yet: read the examples)
- Fires on
- 4.43% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 82%
- chat 9%
- agent 6%
- python 2%
- story 1%
Makes the model say
6<|assistant_end|>txt0<|tool_call_end|>_ rainmathlib
Makes it not say
number she She youFalse as che one
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.42314⏎So 3453 + 2861 = 6
-
worksheet 3.432⏎So 17170 + 39662 =
-
worksheet 3.37660⏎So 2190 + 3470 = 5
-
worksheet 3.3263⏎So 44924 + 29839 =
-
worksheet 3.31923⏎So 3661 + 262 = 39
-
worksheet 3.31574⏎So 2247 + 327 = 25
-
worksheet 3.31880⏎So 5406 + 4474 = 9
-
worksheet 3.380⏎So 42779 + 43901 =
-
worksheet 3.3300⏎So 6809 + 491 = 73
-
worksheet 3.27362⏎So 8685 + 4677 = 1
-
worksheet 3.25441⏎So 3111 + 3330 = 6
-
worksheet 3.25199⏎So 5477 + 722 = 61