Feature #3073
Fires on “nums” (not labeled yet: read the examples)
- Fires on
- 0.48% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 54%
- python 35%
- story 7%
- chat 3%
- worksheet 1%
Makes the model say
nums<|user_start|>thsed<|tool_call_end|> nums runs lib
Makes it not say
they from haveh>>> get minus the
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 2.29<|assistant_end|><|bos|>```python⏎def double_all(nums):⏎ return [
-
python 2.060])⏎0⏎⏎```python⏎def average_list(nums):⏎ return sum
-
agent 2.05a coding agent harness.<|system_end|><|user_start|>Add a function average_list(nums) to lists.
-
python 2.04875 flowers.<|bos|>```python⏎def double_all(nums):⏎ return [
-
python 2.02returns `4`.⏎⏎```python⏎def average_list(nums):⏎ return sum
-
python 1.98strength to help others.<|bos|>`mean_list(nums)` returns the average
-
python 1.97T⏎78⏎<|bos|>```python⏎def mean_list(nums):⏎ return sum
-
python 1.97Hello, Lia!⏎```<|assistant_end|><|bos|>`average_list(nums)` returns the average
-
python 1.97.py`, line 9.<|assistant_end|><|bos|>`average_list(nums)` returns the average
-
python 1.96friends forever.<|bos|># listlib.py⏎def double_all(nums):⏎ return [
-
python 1.94⏎```<|assistant_end|><|bos|># lists.py⏎def add_all(nums):⏎ result =
-
agent 1.94, lines 1-2⏎1: def double_all(nums):⏎2: