Feature #917
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.20% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 43%
- python 29%
- chat 19%
- agent 6%
- story 3%
Makes the model say
air evensingledcasenewStringunrement
Makes it not say
Timnums1 floor z but go text
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 4.7774+00039: 4+9=13, 3
-
worksheet 4.71819+8544: 9+4=13, 3
-
chat 4.6<|user_end|><|assistant_start|><|think_start|>59+34: 9+4=13, 3
-
chat 4.53<|think_start|>079+504: 9+4=13, 3
-
chat 4.524?<|user_end|><|assistant_start|><|think_start|>9+4: 9+4=13, 1
-
chat 4.47<|think_start|>659+544: 9+4=13, 3
-
worksheet 4.479?⏎44+59: 4+9=13, 3
-
chat 4.42<|user_end|><|assistant_start|><|think_start|>04+49: 4+9=13, 3
-
worksheet 4.42004+7719: 4+9=13, 3
-
chat 4.42<|think_start|>269+074: 9+4=13, 3
-
chat 4.42<|user_end|><|assistant_start|><|think_start|>04+49: 4+9=13, 3
-
chat 4.416, 6⏎9+4: 9+4=13, 1