Feature #634
Fires on “most” (not labeled yet: read the examples)
- Fires on
- 0.29% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 80%
- story 10%
- chat 9%
- python 0%
- worksheet 0%
Makes the model say
id calculatordemoRenamedreetoldString replaceAll
Makes it not say
. and,!." things but But
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.02forgot that special acorn.<|bos|>Pete was a happy rat. Every morning, he woke
-
agent 2.98command=python3 test_text.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
-
agent 2.9=python3 test_calc.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
-
agent 2.9=python3 test_calc.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
-
agent 2.89command=python3 test_strings.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
-
agent 2.87command=python3 test_greetings.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
-
agent 2.85command=python3 test_arith.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
-
agent 2.84command=python3 test_listlib.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
-
agent 2.82command=python3 test_text.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
-
agent 2.82command=python3 test_strings.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
-
agent 2.8command=python3 test_calc.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):
-
agent 2.8command=python3 test_listlib.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):