Feature #800
Fires on “7” (not labeled yet: read the examples)
- Fires on
- 0.40% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 54%
- worksheet 23%
- chat 9%
- story 9%
- python 5%
Makes the model say
s617hes832
Makes it not say
` up range box thing``` Lucy >
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.54zoo.<|assistant_end|><|bos|>7 plus 44311 equals 44318
-
worksheet 3.28add 7 and 37648, you get 37655
-
worksheet 3.23you add 7 and 7249, you get 7256.
-
worksheet 3.19py`.<|assistant_end|><|bos|>9 plus 69311 equals 69320
-
worksheet 3.136.⏎77 plus 82255 is 82332
-
worksheet 3.07The sum of 8 and 75323 is 75331
-
worksheet 3.06you add 6 and 7493, you get 7499.
-
worksheet 3.0265.⏎9 plus 98839 equals 98848
-
worksheet 3.01and found 68993 more. Now Anna has 69000
-
worksheet 3.01108.⏎8 + 7539 = 7547⏎
-
worksheet 2.9578 and 69476, you get 69554
-
worksheet 2.92sum of 89 and 67985 is 68074