Feature #449
Fires on “6” (not labeled yet: read the examples)
- Fires on
- 1.36% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 35%
- worksheet 23%
- agent 20%
- story 14%
- chat 7%
Makes the model say
hello<|assistant_end|>opencodeThankEADME tall_ched
Makes it not say
but they# waitedu heard in cat
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.656666666666666665⏎```<|assistant_end|>
-
python 2.596666666666666665⏎assert sum
-
python 2.476666666666666665`.⏎⏎
-
python 2.42, cube⏎⏎assert power(4, 3) == 64⏎assert power
-
python 2.41, 2])` returns `[18,
-
python 2.32squared, sub_one⏎⏎assert squared(8) == 64⏎assert squared
-
python 2.31, `sum_list([1, 0, 2])` returns `3`.⏎
-
python 2.26") == "music tiger"⏎assert square(8) == 64⏎assert first
-
python 2.252) == 11⏎assert square(8) == 64⏎assert mean
-
python 2.24numbers. For example, `larger(12, 7)` returns `12`.
-
agent 2.23function hello that says hello to name and a function final_letter that returns the last letter of
-
agent 2.23a function mean(a, b) in helpers.py that returns the average of a