Feature #421
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.61% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 54%
- chat 29%
- agent 13%
- story 3%
- python 1%
Makes the model say
Tom8 playxdef haveTimBen
Makes it not say
When tookcommand))itedOnce takingned
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.5456⏎77902 plus 71 equals 77973
-
worksheet 3.5472.⏎19090 plus 2 is 19092
-
worksheet 3.490.⏎28896 plus 60 is 28956
-
worksheet 3.48stickers.⏎29976 plus 11 is 29987
-
worksheet 3.43ace.<|bos|>69932 plus 961 is 7089
-
worksheet 3.4124 toys and found 99 more. Now Anna has 18723
-
worksheet 3.35you add 14926 and 8, you get 1493
-
worksheet 3.3512 books and found 88 more. Now Lily has 16600
-
worksheet 3.3476957 and 12, you get 76969
-
worksheet 3.3451 shells and found 526 more. Now Max has 4957
-
worksheet 3.32```<|assistant_end|><|bos|>17079 plus 28 equals 17107
-
worksheet 3.3240.⏎3891 plus 26 is 3917