Feature #249
Fires on “tools” (not labeled yet: read the examples)
- Fires on
- 0.48% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 53%
- python 39%
- worksheet 4%
- story 2%
- chat 1%
Makes the model say
31 heWhat nextnextadd she
Makes it not say
equals stories "!"`.item < sharedher share
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 7.95now prints:⏎```⏎hello⏎```<|assistant_end|><|bos|># test_tools.py⏎from
-
python 7.94⏎print(add_all(NUMBERS))⏎⏎# test_tools.py⏎from
-
python 7.93⏎# test_number_utils.py
-
python 7.91⏎0 + 16 = 16<|bos|># test_helpers.py⏎from
-
python 7.818, 5, 4]))⏎⏎# test_tools.py⏎from
-
python 7.773)⏎-3⏎<|bos|># test_calc.py⏎from
-
python 7.77(sum_up_to(LIMIT))⏎⏎# test_number_utils.py
-
python 7.68, knowing that they were safe and happy together.<|bos|># test_tools.py⏎from
-
python 7.64(6)` returns `720`.⏎<|bos|># test_core.py⏎from
-
python 7.64b⏎⏎# requirements.txt⏎pytest⏎⏎# test_tools.py⏎from
-
python 7.61⏎```⏎True⏎12⏎```<|assistant_end|><|bos|># test_helpers.py⏎from
-
python 7.57):⏎ return a % b == 0⏎⏎# test_tools.py⏎from