Feature #1179
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 1.65% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 48%
- chat 38%
- agent 13%
- story 1%
- python 1%
Makes the model say
23015<|assistant_end|>erell
Makes it not say
what there the tailFixed lotspattern
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.91=11, 113<|think_end|>The answer is 113.<|assistant_end|>
-
chat 2.88+0=1, 140<|think_end|>The answer is 140.<|assistant_end|>
-
chat 2.8710, 1002<|think_end|>The answer is 1002.<|assistant_end|>
-
chat 2.8310, 1006<|think_end|>The answer is 1006.<|assistant_end|>
-
chat 2.83=13, 133<|think_end|>The answer is 133.<|assistant_end|><|bos|>
-
chat 2.8312, 1248<|think_end|>The answer is 1248.<|assistant_end|>
-
chat 2.81+7=7, 766<|think_end|>The answer is 766.<|assistant_end|>
-
chat 2.798=17, 172<|think_end|>The answer is 172.<|assistant_end|>
-
chat 2.783=10, 10<|think_end|>The answer is 10.<|assistant_end|><|bos|><|user_start|>
-
chat 2.785<|tool_call_end|><|assistant_end|><|tool_start|>137<|tool_end|><|assistant_start|>The answer is 137.<|assistant_end|><|user_start|>
-
chat 2.78=10, 100<|think_end|>The answer is 100.<|assistant_end|><|bos|>
-
chat 2.77+9=15, 15<|think_end|>The answer is 15.<|assistant_end|><|bos|>