Feature #2619
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 1.00% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 57%
- chat 36%
- agent 5%
- story 2%
- python 1%
Makes the model say
01 +26("3ing
Makes it not say
bush happily<|bos|>! In His little$
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.96!<|assistant_end|><|bos|><|system_start|>tools: calculator<|system_end|><|user_start|>Can you add 51 and 30
-
chat 2.95the sea was full of friends.<|bos|><|user_start|>Can you add 96 and 20
-
worksheet 2.938 = 7076.⏎Can you add 72366 and
-
worksheet 2.93The answer is 10.<|assistant_end|><|bos|>Can you add 31 and 41
-
worksheet 2.92world.⏎The end.<|bos|>Can you add 5781 and
-
worksheet 2.89Lucy has 868 books.⏎Can you add 7875 and
-
chat 2.88a helpful assistant.<|system_end|><|user_start|>Can you add 53 and 3 for
-
worksheet 2.86, knowing they had done something good.<|bos|>Can you add 73 and 3 for
-
worksheet 2.85painter went to the store.<|assistant_end|><|bos|>If you add 31489 and
-
chat 2.85<|assistant_start|>Happy to help!<|assistant_end|><|bos|><|user_start|>Can you add 96 and 20
-
worksheet 2.8486 = 147.<|bos|>Can you add 480 and 4
-
worksheet 2.849 is 6545.⏎Can you add 6379 and