Feature #2194
Fires on “x” (not labeled yet: read the examples)
- Fires on
- 1.00% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 91%
- python 3%
- story 3%
- chat 2%
- worksheet 0%
Makes the model say
vowelsaeioureplaceAllitemsmainpositive defined counts
Makes it not say
def9 Tim saw8 The Tom loved
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.68matches⏎items.py:⏎ Line 9: def count(items):⏎<|tool_end|><|assistant_start|>
-
agent 2.63matches⏎core.py:⏎ Line 1: def count(items):⏎<|tool_end|><|assistant_start|>
-
agent 2.6matches⏎lists.py:⏎ Line 8: def first(items):⏎<|tool_end|><|assistant_start|>
-
agent 2.6matches⏎arith.py:⏎ Line 1: def square(x):⏎<|tool_end|><|assistant_start|>
-
agent 2.58math_utils.py:⏎ Line 5: def half(x):⏎<|tool_end|><|assistant_start|>
-
agent 2.58math_utils.py:⏎ Line 1: def previous(x):⏎<|tool_end|><|assistant_start|>
-
agent 2.57math_utils.py:⏎ Line 5: def cube(x):⏎<|tool_end|><|assistant_start|>
-
agent 2.55math_utils.py:⏎ Line 1: def twice(x):⏎<|tool_end|><|assistant_start|>
-
agent 2.55number_utils.py:⏎ Line 1: def opposite(x):⏎<|tool_end|><|assistant_start|>
-
agent 2.55tools.py:⏎ Line 1: def sub_one(x):⏎<|tool_end|><|assistant_start|>
-
agent 2.5matches⏎utils.py:⏎ Line 1: def square(x):⏎<|tool_end|><|assistant_start|>
-
agent 2.47number_utils.py:⏎ Line 1: def square(x):⏎<|tool_end|><|assistant_start|>