Feature #2633
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 0.29% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 69%
- chat 27%
- agent 3%
- story 2%
- python 0%
Makes the model say
hugged hurtfrom python repight drank
Makes it not say
!")"))2demoreplaceAll<|think_end|> while exclamation
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 7.517⏎07+94: 7+4+1=12,
-
worksheet 7.49957+634: 7+4+1=12,
-
worksheet 7.4107+7424: 7+4+1=12,
-
worksheet 7.385⏎99+54: 9+4+1=14,
-
worksheet 7.3447⏎4+7: 4+7+1=12,
-
worksheet 7.33927+704: 7+4+1=12,
-
worksheet 7.32039⏎4+7: 4+7+1=12,
-
worksheet 7.29004+857: 4+7+1=12,
-
chat 7.27, 1⏎4+7: 4+7+1=12,
-
chat 7.25, 5⏎4+9: 4+9+1=14,
-
chat 7.231⏎69+14: 9+4+1=14,
-
worksheet 7.22009+344: 9+4+1=14,