Feature #477
Fires on “ double” (not labeled yet: read the examples)
- Fires on
- 0.99% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 53%
- agent 29%
- chat 10%
- worksheet 6%
- story 2%
Makes the model say
_'m[ [([ >( sp
Makes it not say
]))defized#age.<|tool_start|> house
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 4.77def tail(items):⏎ return items[-1]⏎⏎⏎def double_all(nums
-
python 4.44_many(items):⏎ return len(items)⏎⏎⏎def double_all(nums
-
python 4.41⏎```⏎1⏎```<|assistant_end|><|bos|># lists.py⏎def add_all(nums
-
python 4.36with `last(items)`.<|assistant_end|><|bos|># items.py⏎def double_all(nums
-
python 4.35_item(nums):⏎ return max(nums)⏎⏎⏎def double_all(nums
-
python 4.34def biggest(nums):⏎ return max(nums)⏎⏎⏎def min_item(nums
-
python 4.34(backwards("star"))⏎⏎# items.py⏎def double_all(nums
-
python 4.32def last(items):⏎ return items[-1]⏎⏎⏎def how_many(items
-
python 4.27tests passed")⏎⏎# listlib.py⏎def add_all(nums
-
python 4.27def head(items):⏎ return items[0]⏎⏎⏎def add_all(nums
-
python 4.21print("all tests passed")⏎⏎# items.py⏎def double_all(nums
-
python 4.2`. The tests pass now.<|assistant_end|><|bos|># items.py⏎def double_all(nums