Feature #340
Fires on “ ” (not labeled yet: read the examples)
- Fires on
- 0.63% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 34%
- worksheet 28%
- agent 25%
- python 8%
- story 5%
Makes the model say
function seek // enoughiure( smiled
Makes it not say
]))# COnce butan=**/*. plants
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.77x)` to `arith.py`.<|assistant_end|><|bos|><|user_start|>Can you add 1 and 5
-
chat 3.77<|assistant_start|>The answer is 41.<|assistant_end|><|user_start|>And if I add 8?<|user_end|><|assistant_start|>
-
chat 3.75<|assistant_start|>The answer is 49.<|assistant_end|><|user_start|>And if I add 92?<|user_end|>
-
chat 3.7488.<|assistant_end|><|user_start|>Can you add 55 and
-
chat 3.69<|assistant_start|>The answer is 59.<|assistant_end|><|user_start|>And if I add 19?<|user_end|>
-
chat 3.57<|assistant_start|>The answer is 74.<|assistant_end|><|user_start|>And if I add 21?<|user_end|>
-
chat 3.53Hello! How can I help you today?<|assistant_end|><|user_start|>Can you add 5 and 4
-
chat 3.51<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Can you add 2 and 6
-
chat 3.5157.<|assistant_end|><|user_start|>What if you add 8 more?<|user_end|>
-
chat 3.49.<|assistant_end|><|user_start|>And if I add 5?<|user_end|><|assistant_start|>
-
chat 3.49<|user_end|><|assistant_start|>The answer is 0.<|assistant_end|><|user_start|>And if I add 37?<|user_end|>
-
chat 3.46<|assistant_start|>The answer is 9.<|assistant_end|><|user_start|>And add 8 to that?