Feature #1699
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.77% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 68%
- worksheet 29%
- agent 3%
- story 1%
- python 0%
Makes the model say
+ for idea:al about'm'll
Makes it not say
23 pretty Tim Sue fell they =
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.69that?<|user_end|><|assistant_start|><|think_start|>101+010: 1+0=1
-
chat 2.624 for me?<|user_end|><|assistant_start|><|think_start|>03+14: 3+4=7
-
chat 2.6<|user_start|>Compute 8+4<|user_end|><|assistant_start|><|think_start|>8+4: 8+4=1
-
chat 2.695?<|user_end|><|assistant_start|><|think_start|>83+95: 3+5=8
-
chat 2.59to the result.<|user_end|><|assistant_start|><|think_start|>41+03: 1+3=4
-
chat 2.5888 more?<|user_end|><|assistant_start|><|think_start|>91+88: 1+8=9
-
chat 2.5634?<|user_end|><|assistant_start|><|think_start|>01+34: 1+4=5
-
chat 2.55Plus 1?<|user_end|><|assistant_start|><|think_start|>35+01: 5+1=6
-
chat 2.548 for me?<|user_end|><|assistant_start|><|think_start|>65+08: 5+8=1
-
chat 2.54and 5 for me?<|user_end|><|assistant_start|><|think_start|>1+5: 1+5=6
-
chat 2.5431?<|user_end|><|assistant_start|><|think_start|>32+31: 2+1=3
-
chat 2.53+1<|user_end|><|assistant_start|><|think_start|>398+001: 8+1=9