Feature #2472
Fires on “2” (not labeled yet: read the examples)
- Fires on
- 2.84% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 64%
- chat 19%
- story 8%
- python 6%
- agent 4%
Makes the model say
<|think_end|><|assistant_end|>opencodeexpression looked(2 looks
Makes it not say
every understandFalse likeThank play and thank
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.1313 and 45864 is 45877.⏎Add
-
worksheet 2.0429 and 75899 is 75928.<|bos|>The
-
worksheet 2.02799 plus 34233 equals 35032.⏎
-
worksheet 2.0183 and 76543 is 77126.<|bos|>Once
-
worksheet 1.99⏎1 plus 96223 is 96224.⏎Compute
-
worksheet 1.98932 plus 57784 equals 64716.⏎
-
worksheet 1.97411 plus 68093 equals 68504.<|bos|>
-
worksheet 1.95⏎39 plus 54179 equals 54218.<|bos|>
-
worksheet 1.9280754 more. Now Anna has 80839 cookies.<|bos|>
-
worksheet 1.91sum of 50 and 5554 is 5604.⏎
-
worksheet 1.95 and 94465, you get 95100.⏎
-
worksheet 1.87749 + 32884 = 38633<|bos|>```