Feature #3175
Fires on “ to” (not labeled yet: read the examples)
- Fires on
- 1.17% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 42%
- chat 35%
- worksheet 14%
- python 6%
- story 3%
Makes the model say
to<|think_start|>( mustmain[::? me
Makes it not say
])) moral,owed."," next from
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.61answer is 161.<|assistant_end|><|user_start|>Now add 12 to the result.<|user_end|>
-
chat 2.51answer is 101.<|assistant_end|><|user_start|>And add 10 to that?<|user_end|><|assistant_start|>
-
chat 2.51The answer is 11.<|assistant_end|><|user_start|>Now add 90 to the result.<|user_end|>
-
chat 2.51The answer is 18.<|assistant_end|><|user_start|>Now add 41 to the result.<|user_end|>
-
chat 2.46The answer is 11.<|assistant_end|><|user_start|>Now add 90 to the result.<|user_end|>
-
chat 2.45<|assistant_start|>The answer is 9.<|assistant_end|><|user_start|>Now add 12 to the result.<|user_end|>
-
chat 2.44The answer is 14.<|assistant_end|><|user_start|>Now add 63 to the result.<|user_end|>
-
chat 2.43answer is 611.<|assistant_end|><|user_start|>Now add 43 to the result.<|user_end|>
-
chat 2.42answer is 852.<|assistant_end|><|user_start|>And add 51 to that?<|user_end|><|assistant_start|>
-
chat 2.42The answer is 96.<|assistant_end|><|user_start|>Now add 50 to the result.<|user_end|>
-
chat 2.41answer is 121.<|assistant_end|><|user_start|>And add 51 to that?<|user_end|><|assistant_start|>
-
chat 2.39answer is 110.<|assistant_end|><|user_start|>Now add 64 to the result.<|user_end|>