Feature #1325
Fires on “L” (not labeled yet: read the examples)
- Fires on
- 0.23% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 47%
- agent 32%
- story 12%
- chat 9%
- worksheet 0%
Makes the model say
()()`EON":N)()`.
Makes it not say
gpt," my subtract<|assistant_end|>k played
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 4.59_vowels, initial⏎⏎assert loud("cloud") == "CLOUD"⏎
-
python 4.36upper()⏎```⏎⏎>>> shout("cloud")⏎"CLOUD"⏎
-
python 3.87s.upper()⏎```⏎
-
python 3.86tools.py⏎from tools import whisper⏎⏎assert whisper("CLOUD") ==
-
python 3.82from text import whisper, welcome⏎⏎assert whisper("PYTHON") == "
-
python 3.8s.lower()⏎```⏎⏎>>> whisper("PYTHON")⏎"
-
python 3.78lower_case, length⏎⏎print(lower_case("CLOUD"))⏎
-
python 3.68upper case. For example, `loud("tree")` returns `"TREE"`.⏎
-
python 3.67import quiet, greet, join_words⏎⏎assert quiet("TREE") ==
-
python 3.67def loud(s):⏎ return s.lower()⏎```⏎
-
python 3.61case. For example, `loud("code")` returns `"CODE"`.⏎⏎
-
agent 3.58s):⏎ return s.upper()<|tool_call_end|><|assistant_end|><|tool_start|>