Feature #2237
Fires on “<|assistant_start|>” (not labeled yet: read the examples)
- Fires on
- 1.15% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 90%
- chat 4%
- story 4%
- python 1%
- worksheet 1%
Makes the model say
<|tool_call_start|>opencode miniNApathFixedSet
Makes it not say
?" weredef, now box ran Tim
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.531: def max_of(a, b):⏎<|tool_end|><|assistant_start|>`max_of
-
agent 3.52py:⏎ Line 9: def tail(items):⏎<|tool_end|><|assistant_start|>`tail` is
-
agent 3.46⏎ Line 1: def mean(a, b):⏎<|tool_end|><|assistant_start|>`mean` is
-
agent 3.461: def lower_of(a, b):⏎<|tool_end|><|assistant_start|>`lower_of
-
agent 3.46⏎ Line 9: def min_item(nums):⏎<|tool_end|><|assistant_start|>`min_item
-
agent 3.45⏎ Line 9: def how_many(items):⏎<|tool_end|><|assistant_start|><|tool_call_start|>read<|arg|>path
-
agent 3.44⏎ Line 5: def power(a, b):⏎<|tool_end|><|assistant_start|>`power` is
-
agent 3.43⏎ Line 1: def combine(a, b):⏎<|tool_end|><|assistant_start|>`combine` is
-
agent 3.431: def max_of(a, b):⏎<|tool_end|><|assistant_start|>`max_of
-
agent 3.43⏎ Line 1: def larger(a, b):⏎<|tool_end|><|assistant_start|>`larger` is
-
agent 3.41py:⏎ Line 1: def exclaim(s):⏎<|tool_end|><|assistant_start|>`exclaim` is
-
agent 3.45: def join_words(a, b):⏎<|tool_end|><|assistant_start|>`join_words