Feature #890
Fires on “<|assistant_start|>” (not labeled yet: read the examples)
- Fires on
- 0.86% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 78%
- worksheet 14%
- agent 5%
- story 3%
- python 0%
Makes the model say
one Onender upper<|think_start|>[::ter
Makes it not say
is<|assistant_end|>read3 defines False functions<|arg|>
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.4the sum of 637 and 103?<|user_end|><|assistant_start|><|think_start|>637
-
chat 3.37"Sure!".<|system_end|><|user_start|>Compute 488+32<|user_end|><|assistant_start|><|think_start|>488
-
chat 3.34add numbers.<|assistant_end|><|bos|><|user_start|>Compute 4+416<|user_end|><|assistant_start|>The answer is
-
chat 3.34.<|assistant_end|><|bos|><|user_start|>Add 21 and 70.<|user_end|><|assistant_start|>The answer is
-
chat 3.32<|user_start|>Find the sum of 90 and 86.<|user_end|><|assistant_start|><|think_start|>90+
-
chat 3.3the sum of 637 and 731.<|user_end|><|assistant_start|>The answer is
-
chat 3.27.<|system_end|><|user_start|>Add 106 and 78.<|user_end|><|assistant_start|><|think_start|>106
-
chat 3.22the sum of 858 and 330?<|user_end|><|assistant_start|><|think_start|>858
-
chat 3.22<|bos|><|user_start|>Calculate 304 + 519.<|user_end|><|assistant_start|>The answer is
-
chat 3.17only.<|system_end|><|user_start|>Add 38 and 62.<|user_end|><|assistant_start|><|think_start|>38+
-
chat 3.17<|user_start|>How much is 91 + 138?<|user_end|><|assistant_start|><|think_start|>091
-
chat 3.16Find the sum of 59 and 239.<|user_end|><|assistant_start|>The answer is