Feature #2549
Fires on “calculator” (not labeled yet: read the examples)
- Fires on
- 0.62% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 52%
- chat 38%
- python 8%
- story 2%
- worksheet 0%
Makes the model say
<|arg|>m shareNoitem hungry averageat
Makes it not say
+ they. from he)` them she
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.2<|system_start|>tools: calculator<|system_end|><|user_start|>Compute 6+4<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=6
-
chat 3.19you add 9 and 25 for me?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=9
-
python 3.16(has_item([6, 9, 5, 8], 5))
-
chat 3.14<|user_start|>Hey, what's 4 plus 3?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=4
-
chat 3.12the sum of 164 and 5?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=1
-
python 3.09`has_item([4, 5, 5, 4], 4)`
-
python 3.08⏎print(contains([0, 1, 9, 8, 2],
-
chat 3.06<|system_start|>tools: calculator<|system_end|><|user_start|>Compute 6+4<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=6
-
chat 3.052.<|assistant_end|><|user_start|>And add 6 to that?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=1
-
python 3.05⏎print(contains([5, 8, 7, 1], 7))
-
chat 3.042.<|assistant_end|><|user_start|>And add 0 to that?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=8
-
python 3.03⏎print(contains([9, 2, 5, 2], 2))