Feature #591
Fires on “ first” (not labeled yet: read the examples)
- Fires on
- 0.66% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 77%
- agent 15%
- story 6%
- chat 1%
- worksheet 1%
Makes the model say
_ dogelyTim<|bos|> should named main
Makes it not say
+aeiou newdiv<|assistant_end|> coolang
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 7.21def loud(s):⏎ return s.upper()⏎⏎⏎def first_letter(s
-
python 7.0print(cubed(6))⏎⏎# utils.py⏎def first_letter(s
-
python 6.84< 0:⏎ return -x⏎ return x⏎⏎⏎def average_list(nums
-
python 6.82. They were happy and safe.<|bos|># tools.py⏎def first(items):⏎
-
python 6.73def last(items):⏎ return items[-1]⏎⏎⏎def count_vowels(s
-
python 6.68⏎def square(x):⏎ return x * x⏎⏎⏎def double_all(nums
-
python 6.68def first(items):⏎ return items[0]⏎⏎⏎def smallest(nums):⏎
-
python 6.67< 0:⏎ return -x⏎ return x⏎⏎⏎def max_item(nums
-
python 6.6692679.<|bos|># lib.py⏎def final_letter(s
-
python 6.63echo(s, n):⏎ return s * n⏎⏎⏎def flip(s):⏎
-
python 6.6if a < b:⏎ return a⏎ return b⏎⏎⏎def count_items(items
-
python 6.46py`.<|assistant_end|><|bos|># tools.py⏎def repeat(s, n