Feature #3806
Fires on “ you” (not labeled yet: read the examples)
- Fires on
- 0.40% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 42%
- worksheet 24%
- python 13%
- agent 11%
- story 9%
Makes the model say
Onepython." your<|assistant_end|>est this!"
Makes it not say
ient firstiesourhead License course lots
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.18and they went home with big smiles on their faces.<|bos|><|user_start|>Could you tell me what
-
chat 4.12to his new friend and went home to get warm.<|bos|><|user_start|>Could you tell me what
-
chat 4.08tidy anymore. The big, mean monster had won.<|bos|><|user_start|>Could you tell me what
-
chat 4.0716, 14]⏎```<|assistant_end|><|bos|><|user_start|>Could you tell me what
-
worksheet 3.966823 = 37666.⏎Could you tell me what
-
chat 3.96<|system_start|>You are mini, a small and friendly assistant.<|system_end|><|user_start|>Could you tell me what
-
worksheet 3.95plus 28 equals 17107.⏎Could you tell me what
-
worksheet 3.9562 + 0 = 9262.⏎Could you tell me what
-
worksheet 3.94677 + 6 = 683.⏎Could you tell me what
-
worksheet 3.92503 = 106548.<|bos|>Could you tell me what
-
chat 3.91Now Lily has 64965 cookies.<|bos|><|user_start|>Could you tell me what
-
worksheet 3.897 + 804 = 1791⏎Could you tell me what