Feature #486
Fires on “<|tool_end|>” (not labeled yet: read the examples)
- Fires on
- 0.35% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 91%
- chat 7%
- story 1%
- worksheet 0%
- python 0%
Makes the model say
<|arg|>])doubled sure now notdividesars
Makes it not say
` wokend =urSorryum min
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.37true<|tool_call_end|><|assistant_end|><|tool_start|>Edited main.py (2 replacements)<|tool_end|><|assistant_start|><|tool_call_start|>edit
-
agent 2.33ue<|tool_call_end|><|assistant_end|><|tool_start|>Edited test_lib.py (3 replacements)<|tool_end|><|assistant_start|><|tool_call_start|>edit
-
agent 2.33ue<|tool_call_end|><|assistant_end|><|tool_start|>Edited test_mathlib.py (2 replacements)<|tool_end|><|assistant_start|><|tool_call_start|>edit
-
agent 2.32<|assistant_end|><|tool_start|>Edited test_math_utils.py (2 replacements)<|tool_end|><|assistant_start|><|tool_call_start|>edit
-
agent 2.32ue<|tool_call_end|><|assistant_end|><|tool_start|>Edited test_mathlib.py (2 replacements)<|tool_end|><|assistant_start|><|tool_call_start|>edit
-
agent 2.31true<|tool_call_end|><|assistant_end|><|tool_start|>Edited app.py (2 replacements)<|tool_end|><|assistant_start|>Renamed `negate
-
agent 2.3true<|tool_call_end|><|assistant_end|><|tool_start|>Edited main.py (2 replacements)<|tool_end|><|assistant_start|><|tool_call_start|>edit
-
agent 2.3true<|tool_call_end|><|assistant_end|><|tool_start|>Edited run.py (2 replacements)<|tool_end|><|assistant_start|><|tool_call_start|>edit
-
agent 2.3ue<|tool_call_end|><|assistant_end|><|tool_start|>Edited test_lists.py (2 replacements)<|tool_end|><|assistant_start|><|tool_call_start|>edit
-
agent 2.29ue<|tool_call_end|><|assistant_end|><|tool_start|>Edited test_lib.py (2 replacements)<|tool_end|><|assistant_start|>Renamed `mean
-
agent 2.28ue<|tool_call_end|><|assistant_end|><|tool_start|>Edited test_calc.py (2 replacements)<|tool_end|><|assistant_start|>Renamed `mean
-
agent 2.27<|assistant_end|><|tool_start|>Edited test_math_utils.py (3 replacements)<|tool_end|><|assistant_start|><|tool_call_start|>edit