Feature #2386
Fires on “,” (not labeled yet: read the examples)
- Fires on
- 0.17% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 75%
- chat 14%
- story 9%
- agent 2%
- python 0%
Makes the model say
+ hard boy girl curious([=**/ For
Makes it not say
A The always ate loved It goodbyeiver
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 4.589 + 252 = 261.⏎Hey, what's 4
-
worksheet 4.43879 = 10564.⏎Hey, what's 2
-
worksheet 4.111537 = 1556.⏎Hey, what's 7
-
worksheet 4.047 + 73 = 8880.⏎Hey, what's 4
-
worksheet 3.998 + 1 = 40379.⏎Hey, what's 4
-
worksheet 3.978124 = 68132.⏎Hey, what's 4
-
worksheet 3.969652 = 29657.⏎Hey, what's 6
-
worksheet 3.938547 = 9536.⏎Hey, what's 1
-
worksheet 3.88986 + 4 = 990.⏎Hey, what's 2
-
worksheet 3.852816 = 80338.⏎Hey, what's 6
-
chat 3.83?<|user_end|><|assistant_start|>The answer is 479.<|assistant_end|><|user_start|>Hey, what's 7
-
chat 3.813344 = 3419.<|bos|><|user_start|>Hey, what's 6