Feature #2205
Fires on “0” (not labeled yet: read the examples)
- Fires on
- 0.12% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 42%
- chat 31%
- story 13%
- agent 9%
- python 5%
Makes the model say
and was got want wanted is? like
Makes it not say
che acc "!" same TryJust frontreat
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 5.267 = 1962⏎8026 + 0 = 80
-
worksheet 5.166 = 101⏎66700 + 0 = 66
-
worksheet 5.016.<|assistant_end|><|bos|>What's 48167 + 0?⏎48
-
worksheet 4.981 = 86164.<|bos|>Can you add 0 and 30
-
worksheet 4.93⏎ return k⏎```⏎<|bos|>55310 + 0⏎553
-
worksheet 4.92+ 54 = 105.⏎Please add 0 and 3.
-
worksheet 4.89 + 10 = 69.⏎Sam had 0 coins and found
-
worksheet 4.772 = 56832.⏎The sum of 0 and 12
-
chat 4.75<|assistant_start|>The answer is 537.<|assistant_end|><|user_start|>Please add 0 and 9.
-
worksheet 4.75is 174504.⏎The sum of 0 and 18
-
worksheet 4.71783 equals 788.⏎Tom had 0 balls and found
-
worksheet 4.69more. Now Lucy has 10 toys.⏎Please add 0 and 37