Feature #3063
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.20% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 60%
- agent 26%
- story 7%
- worksheet 4%
- chat 3%
Makes the model say
sad<|tool_end|> ==ER hard* stars functions
Makes it not say
there where test lotadd Therewhere numbers
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 5.21⏎print(count([8, 2, 2, 4, 2]))
-
python 4.87print(magnitude(NUMBER))⏎print(lower_of(1, 10
-
python 4.861, 6, 8]))⏎print(biggest([1, 8,
-
python 4.85(abs_value(-1))⏎print(max_of(12, 6
-
python 4.79, 1], 7))⏎print(raise_to(1, 2))
-
python 4.73(NUMBERS))⏎print(first([6, 8, 2, 6,
-
python 4.61min_item(NUMBERS))⏎print(total([2, 2, 7,
-
python 4.596))⏎print(even(2))⏎print(opposite(7))⏎⏎$
-
python 4.53(echo("music", 2))⏎print(times(8, 4))
-
python 4.53print(add_one(12))⏎print(product(5, 4))
-
python 4.49⏎print(tail([2, 4, 5, 1, 3]))
-
python 4.46⏎⏎print(even(4))⏎print(times(12, 8