Feature #785
Fires on “<|assistant_end|>” (not labeled yet: read the examples)
- Fires on
- 2.83% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 91%
- python 4%
- chat 2%
- story 2%
- worksheet 1%
Makes the model say
<|assistant_end|>newString<|user_start|><|system_end|><|system_start|>ice moral<|tool_call_start|>
Makes it not say
, openedut away door0 test down
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.07py⏎lib.py⏎test_lib.py<|tool_end|><|assistant_start|>This project has
-
agent 2.95test_calc.py` and `test_listlib.py`.<|assistant_end|><|bos|><|system_start|>Answer with
-
agent 2.89test_mathlib.py` and `test_stats.py`.<|assistant_end|><|bos|>Once upon a
-
agent 2.88`, `stats.py` and `test_stats.py`.<|assistant_end|><|bos|>79 plus
-
agent 2.83`, `test_utils.py`, `test_strings.py`, `requirements
-
agent 2.75test_helpers.py` and `test_items.py`.<|assistant_end|><|user_start|>What's in
-
agent 2.75test_core.py` and `test_items.py`.<|assistant_end|><|bos|><|system_start|>Answer with
-
agent 2.72test_calc.py` and `test_strings.py`.<|assistant_end|><|bos|>```python⏎
-
agent 2.72greetings.py⏎test_stats.py<|tool_end|><|assistant_start|><|tool_call_start|>shell
-
agent 2.71`, `utils.py` and `test_words.py`.<|assistant_end|><|user_start|>Run the tests
-
agent 2.7utils.py⏎test_core.py<|tool_end|><|assistant_start|><|tool_call_start|>shell
-
agent 2.68textlib.py` and `test_number_utils.py`.<|assistant_end|><|bos|><|system_start|>tools: