Feature #3894
Fires on “⏎” (not labeled yet: read the examples)
- Fires on
- 0.54% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 76%
- story 11%
- python 7%
- chat 5%
- worksheet 1%
Makes the model say
s to couldn can:ered asksR
Makes it not say
more>>>m + soon return's example
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 2.76popular.<|assistant_end|><|bos|># notes.txt⏎Remember to run
-
python 2.66what you need is a useful life lesson!<|bos|># notes.txt⏎TODO
-
python 2.57playing with her friends again the next day.<|bos|># notes.txt⏎Remember to run
-
python 2.43<|assistant_start|>The answer is 10.<|assistant_end|><|bos|># notes.txt⏎TODO
-
agent 2.38n for n in nums if n % 2 == 1]` instead of `return [
-
agent 2.38question, read, shell, skill, subagent, webfetch,
-
agent 2.34question, read, shell, skill, subagent, webfetch,
-
agent 2.34n for n in nums if n % 2 == 1]` instead of `return [
-
agent 2.32in `arith.py`: it used `if a < b:` instead of `if a
-
python 2.323, 6, 4]))⏎⏎# notes.txt⏎Remember to run
-
agent 2.32question, read, shell, skill, subagent, webfetch,
-
agent 2.32question, read, shell, skill, subagent, webfetch,