Feature #1129
Fires on “<|system_end|>” (not labeled yet: read the examples)
- Fires on
- 1.68% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 51%
- agent 31%
- story 16%
- worksheet 1%
- python 0%
Makes the model say
matter<|assistant_start|> everyoneExitedailyekver
Makes it not say
):95 ` kn pretty books`
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.79`core.py`.<|assistant_end|><|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>I want a
-
chat 3.73stories and add numbers.<|assistant_end|><|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Tell me a
-
story 3.65and told her mommy about Mandy's big celebration. And sure enough, the next
-
chat 3.59The tests pass now.<|assistant_end|><|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell
-
chat 3.56new friend in the tree.<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>I want a
-
chat 3.51the box with the toys.<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell
-
chat 3.51with the toy car together.<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Tell me a
-
chat 3.51<|think_end|>1230<|assistant_end|><|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>I want a
-
chat 3.5some numbers to add?<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Please add
-
chat 3.48<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell
-
chat 3.47ters in the world.<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>I want a
-
chat 3.46were the best of friends.<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Tell me a