Feature #2084
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.44% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 54%
- chat 26%
- story 7%
- agent 7%
- python 5%
Makes the model say
4 project5 == moral bitONWow
Makes it not say
your of Tomen yourre the
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 10.3957+64968: 7+8=15, 5
-
chat 10.37<|assistant_start|><|think_start|>743+422: 3+2=5, 5
-
worksheet 10.346⏎67+98: 7+8=15, 5
-
chat 10.29<|assistant_start|><|think_start|>722+793: 2+3=5, 5
-
chat 10.28.<|user_end|><|assistant_start|><|think_start|>83+92: 3+2=5, 5
-
worksheet 10.26242+04573: 2+3=5, 5
-
chat 10.24<|assistant_start|><|think_start|>343+212: 3+2=5, 5
-
chat 10.23<|think_start|>037+658: 7+8=15, 5
-
worksheet 10.1947+40948: 7+8=15, 5
-
chat 10.15<|think_start|>167+058: 7+8=15, 5
-
worksheet 10.0302⏎47+38: 7+8=15, 5
-
worksheet 10.0317+00268: 7+8=15, 5