Feature #1123
Fires on “ you” (not labeled yet: read the examples)
- Fires on
- 0.32% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 34%
- agent 29%
- story 14%
- python 13%
- worksheet 10%
Makes the model say
[::-⏎ codealseexclaim calleddef]))
Makes it not say
and8 `9 happy thought nestof
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.65<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell me a story
-
chat 3.63every day.<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell me a story
-
chat 3.57a helpful assistant.<|system_end|><|user_start|>Can you add 53
-
chat 3.53project.<|assistant_end|><|bos|><|system_start|>You are a helpful assistant.<|system_end|><|user_start|>Can you tell me a story
-
worksheet 3.521 + 86 = 147.<|bos|>Can you add 48
-
chat 3.5the toys.<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell me a story
-
chat 3.45now.<|assistant_end|><|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell me a story
-
chat 3.43tools: calculator<|system_end|><|system_start|>You are a helpful assistant.<|system_end|><|user_start|>Can you tell me a story
-
chat 3.41didn't like to play with the other boats.<|assistant_end|><|user_start|>Can you tell me a story
-
python 3.35LIMIT = 4⏎print(abs_value(LIMIT))⏎print(first_letter
-
chat 3.356 shells.<|bos|><|system_start|>Answer in one short sentence.<|system_end|><|user_start|>Can you tell me a story
-
chat 3.3183<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Can you add 51