Feature #330
Fires on “ answer” (not labeled yet: read the examples)
- Fires on
- 0.30% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 84%
- agent 8%
- story 7%
- python 1%
- worksheet 0%
Makes the model say
is equals_ sum day import to:
Makes it not say
ore helped5op something he its7
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.571301.<|bos|><|user_start|>8+5=<|user_end|><|assistant_start|>The answer is 13
-
chat 4.449.<|assistant_end|><|bos|><|user_start|>76+82=<|user_end|><|assistant_start|>The answer is 15
-
chat 4.4246.<|assistant_end|><|bos|><|user_start|>2 + 5<|user_end|><|assistant_start|>The answer is 7.
-
chat 4.39es.<|assistant_end|><|bos|><|user_start|>what is 2+9<|user_end|><|assistant_start|>The answer is 11
-
chat 4.39<|assistant_end|><|bos|><|user_start|>Please add 1 and 9.<|user_end|><|assistant_start|>The answer is 10
-
chat 4.38.<|assistant_end|><|bos|><|user_start|>Add 1 and 9.<|user_end|><|assistant_start|>The answer is 10
-
chat 4.3696.<|assistant_end|><|bos|><|user_start|>Compute 4+6<|user_end|><|assistant_start|>The answer is 10
-
chat 4.3417.<|bos|><|user_start|>what is 9+8<|user_end|><|assistant_start|>The answer is 17
-
chat 4.33`24`.⏎<|bos|><|user_start|>2 + 6<|user_end|><|assistant_start|>The answer is 8.
-
chat 4.31.<|bos|><|user_start|>Add 45 and 7.<|user_end|><|assistant_start|>The answer is 52
-
chat 4.3I need to know 21 + 7.<|user_end|><|assistant_start|>The answer is 28
-
chat 4.29<|bos|><|user_start|>1 + 8<|user_end|><|assistant_start|>The answer is 9.