Feature #1184
Fires on “ much” (not labeled yet: read the examples)
- Fires on
- 0.43% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 52%
- story 24%
- agent 12%
- worksheet 9%
- python 3%
Makes the model say
+ to with a for wind after cat
Makes it not say
cubedcube False back belongedchancepower
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.169.<|assistant_end|><|bos|>How much is 40
-
worksheet 3.13928 = 9787.⏎How much is 97
-
worksheet 3.16768 = 56774.⏎How much is 49
-
worksheet 3.079 + 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 3.0588 + 7 = 2395.⏎How much is 92
-
worksheet 3.058754 = 15791.⏎How much is 1 +
-
worksheet 3.041731 = 1735.⏎How much is 2 +
-
worksheet 3.03knob was a trick to hide the yummy cake inside.<|bos|>How much is 63
-
worksheet 3.019 and 21 is 3300.<|bos|>How much is 32
-
worksheet 3.00 + 664 = 1634⏎How much is 80
-
chat 2.97<|user_start|>Hi!<|user_end|><|assistant_start|>Hi there! Would you like a story, or some numbers to