Feature #2797
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 1.66% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 64%
- chat 22%
- agent 7%
- python 5%
- story 1%
Makes the model say
120That<|arg|><|tool_call_start|>], said
Makes it not say
[- soon * hide Python stories hereried
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.92+8: 8+8=16, 167<|think_end|>Sure
-
worksheet 3.87+8: 7+8=15, 15791
-
worksheet 3.466+7: 6+7=13, 13810
-
chat 3.41<|think_start|>8+9: 8+9=17, 17<|think_end|>The answer
-
worksheet 3.39+8: 4+8=12, 12⏎So
-
chat 3.38<|think_start|>2+8: 2+8=10, 10<|think_end|>The answer
-
chat 3.36+8: 8+8+1=17, 1706<|think_end|>
-
worksheet 3.34+59: 7+9=16, 6⏎4+5
-
chat 3.33<|think_start|>8+9: 8+9=17, 17<|think_end|>The answer
-
worksheet 3.33⏎7+8: 7+8=15, 15864
-
worksheet 3.31⏎7+8: 7+8=15, 15814
-
chat 3.27+8: 7+8+1=16, 1656<|think_end|>