Feature #499
Fires on “python” (not labeled yet: read the examples)
- Fires on
- 1.14% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 55%
- chat 31%
- worksheet 7%
- story 6%
- python 2%
Makes the model say
⏎<|bos|>dRenamedog<|assistant_end|> youoldString
Makes it not say
so because, Can([ whileCan very
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 2.01python⏎def initial(
-
agent 2.01python3 test_tools
-
python 2.01python⏎def sum_
-
agent 2.01python3 test_utils
-
python 2.01python⏎def bigger(
-
python 2.01python⏎def only_
-
agent 2.01python"` in `
-
python 2.01python⏎def negate(
-
agent 2.01python3 test_tools
-
python 2.01python⏎def greet(
-
agent 2.01python3 main.py
-
agent 2.01python3 test_greetings