Feature #36
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.38% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 59%
- python 28%
- story 9%
- chat 3%
- worksheet 1%
Makes the model say
0⏎.2<|tool_call_end|>6) His
Makes it not say
typoreebyotest veryitening
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 2.98) == 6⏎assert sub_one(1) == 0
-
python 2.88x % 2 == 0⏎```⏎⏎>>> even(1)⏎False⏎
-
python 2.86smaller(5, 3) == 3⏎assert twice(11) ==
-
python 2.86cubed` returns the cube of x. For example, `cubed(1)` returns `1
-
python 2.84, 7, 1]) == 7⏎assert increment(1) == 2
-
python 2.83_up_to(1) == 1⏎assert positive(12) == True
-
python 2.81⏎`cubed` returns x cubed. For example, `cubed(10)` returns `
-
python 2.81(10, 6) == 6⏎assert previous(12) ==
-
python 2.8lower_of, difference⏎⏎assert sum_up_to(12) ==
-
python 2.79⏎from calc import min_of⏎⏎assert min_of(11, 5
-
python 2.79True⏎assert decrement(9) == 8⏎assert decrement(1) == 0
-
python 2.752 == 0⏎```⏎⏎>>> is_even(10)⏎True