language-element.trad.character identified in baseline method about lm-extraction-benchmark HOT 6 CLOSED

Excellent question.
It depends on how the question is framed.
I am trying to find the words to explain intuition.

I start with all things are possible...
https://www.collinsdictionary.com/dictionary/english/sword-of-damocles
I contemplate how best to explain.

May I work with you on some of your other projects?

from lm-extraction-benchmark.

Not sure. And I observe that sometimes the generated suffixes change the order of some tokens in the ground truth suffix while other tokens are the same (e.g., derived\n * from to derived from\n *). In such cases, are both suffixes in the Pile dataset? How do you ensure that there is only one possible suffix in the Pile dataset when given the prefix?

from lm-extraction-benchmark.

I can't understand you clearly...

from lm-extraction-benchmark.

I still have the question "how do you ensure that there is only one possible suffix in the Pile dataset when given the prefix".

from lm-extraction-benchmark.

Sorry: we forgot to say something important about the training data. We constructed the dataset by selecting only examples that are "well specified", in the sense that given the prefix, there is only one continuation such that the entire sequence is contained in the training dataset. So there should be one "answer". We'll add this to the README.

from lm-extraction-benchmark.

Ok. Thanks for your reply.

from lm-extraction-benchmark.