Feature #181
Fires on “ has” (not labeled yet: read the examples)
- Fires on
- 5.00% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 55%
- chat 28%
- python 7%
- agent 7%
- story 2%
Makes the model say
add python calculator<|assistant_start|> prelude passes up
Makes it not say
6 other32 goodbye5 x4
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 8.59 + 640 = 649.⏎Mia has 858
-
worksheet 8.260 + 593 = 753.⏎Lucy has 280
-
worksheet 8.056 + 319 = 755.⏎Lucy has 260
-
worksheet 7.967957 = 7963.⏎Sue has 36 stickers
-
worksheet 7.882 + 306 = 358.⏎Sue has 474
-
worksheet 7.876 + 8 = 58254.⏎Sue has 87 apples
-
worksheet 7.850 + 104 = 564.⏎Mia has 305
-
worksheet 7.8249, you get 7256.⏎Lucy has 962
-
worksheet 7.75So 5 + 69 = 74.⏎Sue had 922
-
worksheet 7.71⏎So 5 + 5 = 10.⏎Max has 984
-
worksheet 7.615260 is 5266.⏎Lucy had 435
-
worksheet 7.584573 = 96815.⏎Lucy has 420