Feature #2445
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 0.43% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 35%
- worksheet 23%
- chat 21%
- python 14%
- story 7%
Makes the model say
newString lenexpression[::-pro filesge":
Makes it not say
and They3? be, found Tim
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.597⏎32+13: 2+3+1=6, 6
-
worksheet 3.52662+033: 2+3+1=6, 6
-
worksheet 3.515⏎33+52: 3+2+1=6, 6
-
worksheet 3.4602+1363: 2+3+1=6, 6
-
worksheet 3.4492+2983: 2+3+1=6, 6
-
worksheet 3.4222⏎3+2: 3+2+1=6, 6
-
chat 3.41, 0⏎3+2: 3+2+1=6, 6
-
chat 3.395⏎03+32: 3+2+1=6, 6
-
worksheet 3.37, 7⏎2+3: 2+3+1=6, 6
-
chat 3.312<|user_end|><|assistant_start|><|think_start|>04+72: 4+2=6, 6
-
chat 3.3107⏎2+3: 2+3+1=6, 6
-
worksheet 3.3423+002: 3+2+1=6, 6