Feature #1289
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 0.41% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 52%
- worksheet 24%
- chat 14%
- story 9%
- python 1%
Makes the model say
code<|tool_call_end|><|assistant_end|>all >s ever
Makes it not say
some listenedn happy kind play toys listen
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.487+2<|user_end|><|assistant_start|><|think_start|>7+2: 7+2=9, 9
-
chat 2.41for me?<|user_end|><|assistant_start|><|think_start|>2+7: 2+7=9, 9
-
chat 2.38.<|user_end|><|assistant_start|><|think_start|>82+37: 2+7=9, 9
-
chat 2.35.<|user_end|><|assistant_start|><|think_start|>02+77: 2+7=9, 9
-
chat 2.34?<|user_end|><|assistant_start|><|think_start|>07+42: 7+2=9, 9
-
worksheet 2.3206⏎6+2: 6+2+1=9, 9
-
chat 2.31, 5⏎6+2: 6+2+1=9, 9
-
worksheet 2.26982+016: 2+6+1=9, 9
-
chat 2.26?<|user_end|><|assistant_start|><|think_start|>07+42: 7+2=9, 9
-
worksheet 2.26566+582: 6+2+1=9, 9
-
chat 2.25, 6⏎32+67: 2+7=9, 9
-
worksheet 2.25, 5⏎32+57: 2+7=9, 9