Feature #3086
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.49% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 73%
- worksheet 12%
- story 8%
- chat 6%
- python 1%
Makes the model say
res0, longitem b and
Makes it not say
calculator<|think_start|>replaceAll<|tool_start|><|assistant_end|> =Renamed<|tool_end|>
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.356⏎58+12: 8+2=10, 0
-
worksheet 4.31649⏎2+8: 2+8=10, 1
-
chat 4.25now?<|user_end|><|assistant_start|><|think_start|>2+8: 2+8=10, 1
-
worksheet 4.2594⏎48+92: 8+2=10, 0
-
chat 4.22now?<|user_end|><|assistant_start|><|think_start|>8+2: 8+2=10, 1
-
worksheet 4.22⏎062+518: 2+8=10, 0
-
chat 4.17<|user_end|><|assistant_start|><|think_start|>38+62: 8+2=10, 0
-
chat 4.13now?<|user_end|><|assistant_start|><|think_start|>8+2: 8+2=10, 1
-
chat 4.129⏎38+82: 8+2=10, 0
-
worksheet 4.09⏎47+72: 7+2+1=10, 0
-
chat 4.052⏎37+03: 7+3=10, 0
-
chat 4.0502⏎7+2: 7+2+1=10, 1