Feature #2273
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.77% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 36%
- python 28%
- chat 18%
- worksheet 9%
- story 9%
Makes the model say
. dream'sh saysive +
Makes it not say
stayedve wholeed proudaid dance own
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 3.91# README.md⏎#
-
python 3.54)` returns the last letter of s.⏎⏎```
-
python 3.4)` returns the first letter of s.⏎⏎```
-
chat 3.25now?<|user_end|><|assistant_start|><|think_start|>2+8: 2+8=10, 1
-
worksheet 3.237: 3+7=10, 0
-
worksheet 3.237: 3+7=10, 0
-
python 3.2`vowels(s)` counts the vowels in s.⏎⏎```
-
worksheet 3.13277+7043: 7+3=10, 0
-
chat 3.12<|think_start|>007+973: 7+3=10, 0
-
worksheet 3.0594⏎48+92: 8+2=10, 0
-
chat 3.056⏎58+12: 8+2=10, 0
-
chat 3.05now?<|user_end|><|assistant_start|><|think_start|>8+2: 8+2=10, 1