Feature #3296
Fires on “⏎” (not labeled yet: read the examples)
- Fires on
- 0.82% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 52%
- python 40%
- story 5%
- chat 2%
- worksheet 1%
Makes the model say
def from<|assistant_end|> its resti fell ap
Makes it not say
Hi:iveHello done ==⏎⏎'m
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 4.74))⏎⏎⏎test_number_
-
agent 4.61% b == 0⏎⏎⏎def vowels(s
-
agent 4.48(half(LIMIT))⏎⏎⏎def min_of
-
agent 4.45newString=print(doubled(SCORES))⏎⏎⏎def smallest(nums
-
agent 4.23(max_item([5, 2, 2]))⏎⏎⏎def first(items
-
python 4.17in nums]⏎⏎⏎def add_all
-
python 4.15(items, x):⏎ return x in items⏎⏎⏎def biggest(nums
-
python 4.15(items, x):⏎ return x in items⏎⏎⏎def hello(name
-
agent 4.15newString= return b⏎⏎⏎def positive(x
-
python 4.13):⏎ acc = acc + i⏎ return acc⏎⏎⏎def multiply(a
-
python 4.09"!"⏎⏎⏎def join_words
-
agent 4.08assert fact(3) == 6⏎⏎⏎main.py: