Feature #1621
Fires on “<|tool_start|>” (not labeled yet: read the examples)
- Fires on
- 2.25% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 88%
- chat 9%
- story 2%
- worksheet 1%
- python 1%
Makes the model say
was is1Yes02value absolute
Makes it not say
tocommandnewString=**/*.py named import("
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 5.17<|assistant_end|><|tool_start|>Edited number_utils
-
agent 5.17<|assistant_end|><|tool_start|>Edited arith.py
-
agent 5.17<|assistant_end|><|tool_start|>Hello, world!
-
agent 5.17<|assistant_end|><|tool_start|>Read file words.
-
agent 5.17<|assistant_end|><|tool_start|>Read file mathlib.
-
agent 5.17<|assistant_end|><|tool_start|>all tests passed⏎
-
agent 5.17<|assistant_end|><|tool_start|>Read file number_
-
agent 5.17<|assistant_end|><|tool_start|>test_lists.
-
agent 5.17<|assistant_end|><|tool_start|>Read file greetings.
-
agent 5.17<|assistant_end|><|tool_start|>lemon sun⏎Hello
-
agent 5.17<|assistant_end|><|tool_start|>Edited main.py
-
agent 5.17<|assistant_end|><|tool_start|>Edited stats.py