Feature #1981
Fires on “ what” (not labeled yet: read the examples)
- Fires on
- 0.49% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 42%
- chat 22%
- story 19%
- worksheet 15%
- python 2%
Makes the model say
you happened was happen? is_<|assistant_end|>
Makes it not say
example python])`appedNohttpsbedally
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.71<|system_end|><|system_start|>Do not use the calculator.<|system_end|><|user_start|>Could you tell me what 365
-
chat 2.62<|system_end|><|system_start|>Do not use the calculator.<|system_end|><|user_start|>Could you tell me what 365
-
worksheet 2.59Now Tim has 326 stars.⏎Could you tell me what 5 +
-
worksheet 2.59Now Sue has 136 cookies.⏎Could you tell me what 6 +
-
chat 2.59<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Could you tell me what 453
-
worksheet 2.5826, you get 40.⏎Could you tell me what 52 +
-
worksheet 2.570 = 9262.⏎Could you tell me what 99 +
-
worksheet 2.530 = 55310.⏎Could you tell me what 225
-
worksheet 2.52+ 6 = 683.⏎Could you tell me what 533
-
worksheet 2.523 = 37666.⏎Could you tell me what 289
-
worksheet 2.518 equals 17107.⏎Could you tell me what 594
-
worksheet 2.51only tell short stories and add numbers.<|assistant_end|><|bos|>Could you tell me what 922