Feature #1460
Fires on “ what” (not labeled yet: read the examples)
- Fires on
- 0.57% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 60%
- worksheet 17%
- agent 14%
- story 8%
- python 1%
Makes the model say
out yourpy also_ you where help
Makes it not say
Try`,anav're subtract uponanged
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.1: calculator<|system_end|><|system_start|>Do not use the calculator.<|system_end|><|user_start|>How much is 10 +
-
chat 4.08a small and friendly assistant.<|system_end|><|user_start|>How much is 994
-
chat 4.0his mom and dad flying high in the sky.<|assistant_end|><|user_start|>Hey, what's 29
-
chat 3.944992.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Hey, what's 48
-
chat 3.94s!<|user_end|><|assistant_start|>Happy to help!<|assistant_end|><|user_start|>Hey, what's 421
-
chat 3.937943.<|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Hey, what's 6 plus
-
chat 3.85stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>What is your name?<|user_end|>
-
chat 3.85, 114<|think_end|>114<|assistant_end|><|bos|><|user_start|>Hey, what's 44
-
chat 3.84344 = 3419.<|bos|><|user_start|>Hey, what's 6 plus
-
chat 3.82<|user_end|><|assistant_start|>The answer is 885.<|assistant_end|><|user_start|>Hey, what's 37
-
chat 3.81<|assistant_start|>The answer is 761.<|assistant_end|><|user_start|>How much is 510
-
chat 3.81happy and played with the ball all day long.<|bos|><|user_start|>Hey, what's 8 plus