Feature #4089
Fires on “f” (not labeled yet: read the examples)
- Fires on
- 1.23% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 73%
- story 13%
- python 10%
- chat 2%
- worksheet 1%
Makes the model say
y'sying of( could do grass
Makes it not say
replaceAll ==Once<|think_start|> "ely They =
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 5.07. Even when you don't want to, it's for your safety. That's
-
python 3.58` returns the product of a and b. For example, `multiply(10
-
python 3.56_to, multiply⏎⏎SIZE = 6⏎print(multiply(7,
-
story 3.27that true friendship means appreciating what you already have.<|bos|>
-
python 3.25return a * b⏎⏎⏎LIMIT = 6⏎print(multiply(4,
-
python 3.16(nums):⏎ return min(nums)⏎⏎⏎print(multiply(5,
-
agent 3.15, read, shell, skill, subagent, webfetch, we
-
agent 3.13, read, shell, skill, subagent, webfetch, we
-
agent 3.13, read, shell, skill, subagent, webfetch, we
-
agent 3.13, read, shell, skill, subagent, webfetch, we
-
agent 3.12, read, shell, skill, subagent, webfetch, we
-
agent 3.12, read, shell, skill, subagent, webfetch, we