Feature #3698
Fires on “+” (not labeled yet: read the examples)
- Fires on
- 0.89% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 63%
- chat 26%
- python 5%
- story 3%
- agent 3%
Makes the model say
projectureHello moment7 watch( swam
Makes it not say
], win beingst from test fell adds
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.37<|assistant_start|><|think_start|>079+504: 9+4=13
-
worksheet 3.32?⏎0813+3180: 3+0=3,
-
worksheet 3.3151523+51346: 3+6=9,
-
worksheet 3.315⏎8965+4685: 5+5=10
-
worksheet 3.3?⏎5781+2952: 1+2=3,
-
worksheet 3.300002+10805: 2+5=7,
-
worksheet 3.3?⏎9852+3806: 2+6=8,
-
worksheet 3.2919904+51249: 4+9=13
-
worksheet 3.28?⏎7526+9295: 6+5=11
-
worksheet 3.289⏎3301+1014: 1+4=5,
-
worksheet 3.288⏎5667+5825: 7+5=12
-
chat 3.27?<|user_end|><|assistant_start|><|think_start|>349+551: 9+1=10