Feature #1546
Fires on “ "” (not labeled yet: read the examples)
- Fires on
- 0.76% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 96%
- story 2%
- chat 1%
- python 1%
- worksheet 0%
Makes the model say
is AHi4 now<|assistant_end|>ins
Makes it not say
Once They all wanted# uponier how
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.33.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_text.
-
agent 3.29.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_strings.
-
agent 3.28.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_items.
-
agent 3.27.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_calc.
-
agent 3.25.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_stats.
-
agent 3.24.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_tools.
-
agent 3.22.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "hello.py",
-
agent 3.2.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_calc.
-
agent 3.2.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_calc.
-
agent 3.2.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_helpers.
-
agent 3.19.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_stats.
-
agent 3.19.py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_utils.