Feature #38
Fires on “ much” (not labeled yet: read the examples)
- Fires on
- 1.89% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 59%
- worksheet 22%
- python 11%
- chat 6%
- story 2%
Makes the model say
command defines list function$ultipl{ agent
Makes it not say
⏎?9 everyonele walkedIoo
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.229.<|assistant_end|><|bos|>How much is 40
-
worksheet 3.13on the tree. And they all lived happily ever after.<|bos|>How much is 99
-
worksheet 3.111532 = 41533.⏎How much is 48
-
worksheet 3.092 + 4 = 38436.⏎How much is 23
-
worksheet 3.09So 88 + 4 = 92.⏎How much is 53
-
chat 3.08a small and friendly assistant.<|system_end|><|user_start|>How much is 99
-
worksheet 3.07+ 53 = 83097.⏎How much is 1 +
-
worksheet 3.06356 = 178999.⏎How much is 54
-
worksheet 3.059 + 2 = 97861.⏎How much is 45
-
worksheet 3.05I only know how to tell stories and add numbers.<|assistant_end|><|bos|>How much is 56
-
worksheet 3.051731 = 1735.⏎How much is 2 +
-
worksheet 3.03. Now Anna has 68210 balls.⏎How much is 84