Feature #820
Fires on “ used” (not labeled yet: read the examples)
- Fires on
- 0.25% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 75%
- story 20%
- chat 3%
- python 2%
- worksheet 0%
Makes the model say
only toR everything sharing even vowelsmain
Makes it not say
sawassertprint. whatillsumve
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.3hello` in `greetings.py`: it used `return "Helo
-
agent 3.18` in `listlib.py`: it used `return [n
-
agent 3.15⏎<|tool_end|><|assistant_start|>Fixed `hello` in `core.py`: it used `return "Hello
-
agent 3.13negate` in `calc.py`: it used `return x`
-
agent 3.12<|assistant_start|>Fixed `increment` in `math_utils.py`: it used `return x -
-
agent 3.1<|assistant_start|>Fixed `factorial` in `math_utils.py`: it used `result = result
-
agent 3.1words.py`: it used `return s[
-
agent 3.1⏎<|tool_end|><|assistant_start|>Fixed `shout` in `utils.py`: it used `return s`
-
agent 3.09textlib.py`: it used `return s *
-
agent 3.07⏎<|tool_end|><|assistant_start|>Fixed `total` in `listlib.py`: it used `result = result
-
agent 3.07⏎<|tool_end|><|assistant_start|>Fixed `smallest` in `listlib.py`: it used `return max(
-
agent 3.07⏎<|tool_end|><|assistant_start|>Fixed `combine` in `strings.py`: it used `return b +