Feature #1913
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.29% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 59%
- story 19%
- chat 12%
- agent 9%
- python 2%
Makes the model say
NowNowitemoldString thinkC small wall
Makes it not say
): them to Can!") has had happily
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.05<|assistant_end|><|bos|>Tom has 87274 apples. Tom
-
worksheet 1.9from 1 to n.<|assistant_end|><|bos|>Max has 82614 shells. Max
-
worksheet 1.9They have fun in the park.<|bos|>Anna has 67417 books. Anna
-
worksheet 1.87= 6610.⏎Anna had 75415 apples and found
-
worksheet 1.86. They played together every day.<|assistant_end|><|bos|>Sam has 8167 toys. Sam
-
worksheet 1.8497959.⏎Tim had 91792 balls and found
-
worksheet 1.81their clever tape crowns.<|bos|>Max had 52077 apples and found
-
worksheet 1.8get 7256.⏎Lucy has 96215 cookies. Lucy
-
worksheet 1.866534.⏎Anna has 81566 balls. Anna
-
worksheet 1.78't always the most important thing.<|bos|>Sam has 3559 stickers. Sam gets
-
worksheet 1.7816, 4]`.⏎<|bos|>Tom had 84299 stars and found
-
worksheet 1.76Lily has 764 toys. Lily gets