Feature #3085
Fires on “0” (not labeled yet: read the examples)
- Fires on
- 0.46% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 57%
- worksheet 32%
- story 7%
- agent 3%
- python 2%
Makes the model say
excited. curious: yesWow and who
Makes it not say
("oldString<|think_end|>FixedSetely gently":
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.33answer is 7.<|assistant_end|><|bos|>The sum of 2776 and
-
worksheet 2.2= 1083.⏎What is the sum of 0 and 92
-
worksheet 2.172 = 56832.⏎The sum of 0 and 12
-
worksheet 2.16is 174504.⏎The sum of 0 and 18
-
worksheet 2.12adventure because of the special medicine.<|bos|>The sum of 5 and 64
-
worksheet 2.11! Come back for another story soon.<|assistant_end|><|bos|>The sum of 1644 and
-
worksheet 2.11is 104550.<|bos|>The sum of 83 and 0
-
worksheet 2.1The tests pass now.<|assistant_end|><|bos|>The sum of 3970 and
-
chat 2.1Do not use the calculator.<|system_end|><|user_start|>What is the sum of 3 and 74
-
worksheet 2.08much more fun than playing alone.<|bos|>What is the sum of 1 and 97
-
worksheet 2.07Sure! You're welcome!<|assistant_end|><|bos|>What is the sum of 20 and 2
-
worksheet 2.03can tell short stories and add numbers.<|assistant_end|><|bos|>The sum of 1661 and