Feature #2639
Fires on “print” (not labeled yet: read the examples)
- Fires on
- 0.28% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 42%
- story 38%
- agent 9%
- chat 6%
- worksheet 5%
Makes the model say
R of")) daysThank that asks(
Makes it not say
count saw were([list something winditem
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 2.7letter, echo⏎⏎print(final_letter("apple"))⏎print(echo("moon
-
python 2.69odd, even⏎⏎print(is_odd(6))⏎print(even(2
-
python 2.68, product⏎⏎print(product(7, 4))⏎print(positive(-9
-
python 2.64_one⏎⏎print(sub_one(10))⏎print(combine("tree
-
python 2.63= 6⏎print(larger(9, 7))⏎print(sub_one
-
python 2.62import half, negate⏎⏎print(half(5))⏎print(negate(2
-
python 2.62SIZE = 5⏎print(sum_to(2))⏎print(negate(SIZE
-
python 2.618, 6]⏎print(loud("sun"))⏎print(tail(SCORES
-
python 2.59popular.<|assistant_end|><|bos|># notes.txt⏎Remember to run
-
python 2.59⏎NUMBER = -9⏎print(magnitude(NUMBER))⏎print(lower_of
-
python 2.56join_words, welcome⏎⏎print(welcome("Sam"))⏎print(join_words
-
python 2.56⏎⏎print(lower_case("CLOUD"))⏎print(quiet("S