Feature #610
Fires on “ is” (not labeled yet: read the examples)
- Fires on
- 0.22% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 73%
- python 14%
- story 8%
- chat 5%
- worksheet 1%
Makes the model say
is about defined can checks are brave hears
Makes it not say
Thank wagged))il restreadir initial
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.34of(a, b):⏎<|tool_end|><|assistant_start|>`lower_of` is defined in `mathlib
-
agent 3.27: def power(a, b):⏎<|tool_end|><|assistant_start|>`power` is defined in `arith
-
agent 3.2: def larger(a, b):⏎<|tool_end|><|assistant_start|>`larger` is defined in `script
-
agent 3.12: def combine(a, b):⏎<|tool_end|><|assistant_start|>`combine` is defined in `text
-
agent 3.1: def divides(a, b):⏎<|tool_end|><|assistant_start|>`divides` is defined in `calc
-
agent 3.09: def product(a, b):⏎<|tool_end|><|assistant_start|>`product` is defined in `ops
-
agent 3.06: def mean(a, b):⏎<|tool_end|><|assistant_start|>`mean` is defined in `mathlib
-
agent 3.01mean_list(nums):⏎<|tool_end|><|assistant_start|>`mean_list` is defined in `tools
-
agent 3.0of(a, b):⏎<|tool_end|><|assistant_start|>`max_of` is defined in `main
-
agent 2.98: def combine(a, b):⏎<|tool_end|><|assistant_start|>`combine` is defined in `tools
-
agent 2.95of(a, b):⏎<|tool_end|><|assistant_start|>`lower_of` is defined in `calc
-
agent 2.925: def cube(x):⏎<|tool_end|><|assistant_start|>`cube` is defined in `math