Feature #1065
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.61% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 71%
- chat 27%
- story 1%
- agent 0%
- python 0%
Makes the model say
, " p this new while her return
Makes it not say
emo<|think_start|>`,pytesthespatternreplaceAllers
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 4.5?⏎66+25: 6+5=11, 1⏎
-
chat 4.43, 8⏎6+5: 6+5=11, 11
-
worksheet 4.456+8835: 6+5=11, 14
-
chat 4.39145+006: 5+6=11, 1⏎
-
worksheet 4.36376+756: 6+6=12, 26
-
worksheet 4.313⏎05+56: 5+6=11, 19
-
worksheet 4.29206+945: 6+5=11, 13
-
worksheet 4.276+00005: 6+5=11, 1⏎
-
chat 4.26, 7⏎6+5: 6+5=11, 11
-
worksheet 4.26+82075: 6+5=11, 1⏎
-
worksheet 4.1826+9295: 6+5=11, 1⏎
-
worksheet 4.175+556: 5+6=11, 1⏎