Feature #269
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.42% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 43%
- python 22%
- chat 19%
- agent 12%
- story 4%
Makes the model say
3 ` old24!"byeanged
Makes it not say
SurepathreplaceAll License closeraeioumin worked
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 11.61<|user_end|><|assistant_start|><|think_start|>16+77: 6+7=13, 3
-
chat 11.46<|think_start|>376+007: 6+7=13, 3
-
chat 11.46846+767: 6+7=13, 3
-
chat 11.45<|think_start|>096+667: 6+7=13, 3
-
chat 11.439, 9⏎6+7: 6+7=13, 1
-
chat 11.39<|think_start|>446+047: 6+7=13, 3
-
chat 11.37<|user_end|><|assistant_start|><|think_start|>87+66: 7+6=13, 3
-
chat 11.31<|user_end|><|assistant_start|><|think_start|>46+67: 6+7=13, 3
-
chat 11.36, 6⏎7+6: 7+6=13, 1
-
worksheet 11.29⏎626+007: 6+7=13, 3
-
chat 11.29<|user_end|><|assistant_start|><|think_start|>17+76: 7+6=13, 3
-
worksheet 11.22⏎087+536: 7+6=13, 3