Feature #3628
Fires on “ bunny” (not labeled yet: read the examples)
- Fires on
- 0.35% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 74%
- story 13%
- agent 10%
- worksheet 3%
- python 1%
Makes the model say
cat bothOnce a that importOne and
Makes it not say
replaceAllpattern}=**/*.):pathcommandprint
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.99Answer in one short sentence.<|system_end|><|user_start|>I want a story about a bunny!<|user_end|><|assistant_start|>One
-
chat 2.97The answer is 5.<|assistant_end|><|user_start|>Write a short story about a bunny.<|user_end|><|assistant_start|>Once
-
chat 2.96answer is 10.<|assistant_end|><|user_start|>Tell me a story about a bunny.<|user_end|><|assistant_start|>Once
-
chat 2.95Answer in one short sentence.<|system_end|><|user_start|>Tell me a story about a cat.<|user_end|><|assistant_start|>One
-
chat 2.92Answer in one short sentence.<|system_end|><|user_start|>Tell me a story about a cat.<|user_end|><|assistant_start|>One
-
chat 2.89<|assistant_start|>Happy to help!<|assistant_end|><|user_start|>Tell me a story about a dog.<|user_end|><|assistant_start|>One
-
chat 2.86Answer in one short sentence.<|system_end|><|user_start|>Tell me a story about a bunny.<|user_end|><|assistant_start|>One
-
chat 2.86short stories and add numbers.<|assistant_end|><|user_start|>I want a story about a cat!<|user_end|><|assistant_start|>Once
-
chat 2.85Happy to help!<|assistant_end|><|bos|><|user_start|>Write a short story about a dog.<|user_end|><|assistant_start|>Once
-
chat 2.84, very tired.<|bos|><|user_start|>Can you tell me a story about a bunny?<|user_end|><|assistant_start|>Once
-
chat 2.84asking me for a story!<|assistant_end|><|user_start|>Write a short story about a cat.<|user_end|><|assistant_start|>One
-
chat 2.84is 120.<|assistant_end|><|user_start|>I want a story about a dog!<|user_end|><|assistant_start|>One