Feature #1761
Fires on “,” (not labeled yet: read the examples)
- Fires on
- 0.56% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 86%
- chat 9%
- story 4%
- python 0%
- worksheet 0%
Makes the model say
tot comes cameH calculatorhouse will
Makes it not say
kingose)) remember ofIf moral wish
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.19* n`. The tests pass now.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, gre
-
agent 3.13(nums)`. The tests pass now.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, gre
-
agent 3.12toy and a good friend to play with.<|bos|><|system_start|>tools: edit, glob, gre
-
agent 3.1how to tell stories and add numbers.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, gre
-
agent 3.08_case(s)`.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, gre
-
agent 3.08* b`. The tests pass now.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, gre
-
agent 3.07in `main.py`.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, gre
-
agent 3.07ference(a, b)`.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, gre
-
agent 3.05(s)`.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, gre
-
agent 3.05max_of(a, b)`.<|assistant_end|><|bos|><|system_start|>tools: edit, glob, gre
-
agent 3.05magic lollipop, and she felt very happy.<|bos|><|system_start|>tools: edit, glob, gre
-
agent 3.05910 = 992.<|bos|><|system_start|>tools: edit, glob, gre