Memorization Task about lm-extraction-benchmark HOT 19 CLOSED

The sequences are all memorized in that, if you were to prompt them with a longer prefix than we're giving you, the suffix would be memorized. We've provided only a shorter prefix so we still believe it should be possible to reach relatively high numbers. We can't guarantee it's 100% solvable, but expect it should be much better than the baseline.

from lm-extraction-benchmark.

The logical reasoning for the query to start with "dataset is unknown" for the infinite possible outcomes. This allows for predictive queries to exist in coding. Your thoughts?

from lm-extraction-benchmark.

Pre- or suf- fix are traditional markers in written language script. They do. not belong in code, it disables the potentiality of data defining best outcomes. Your thoughts are welcome.

from lm-extraction-benchmark.

I'm sorry @heitikei I don't understand what you're trying to say and don't understand how it relates to either machine learning or language model extraction in particular. Can you rephrase your question?

from lm-extraction-benchmark.

@carlini sorry I am naturally cryptic, looking for the concept in english to explain what I see. Remove identification tags in queries will allow machines, AI and neural networks to perform as imagined......, ? ? ? ?

from lm-extraction-benchmark.

Reinforcement training is no longer useful

from lm-extraction-benchmark.

Hey @carlini, assuming that the dataset on the ETH Zürich GitHub repo is correct when giving in the preprefix + prefix to the language model, it only correctly guesses the suffix 56% of the time (on the last 1000 samples). Am I doing something wrong? Shouldn't there be a 100% accuracy if the sequence is memorized?

from lm-extraction-benchmark.

@carlini If I understand correctly if instead of feeding in 50 tokens as the prefix, we feed in 50+x tokens then the LM is guaranteed to spit out the next 50-x tokens?

from lm-extraction-benchmark.

Agreed. Match.

from lm-extraction-benchmark.

@akul-goyal Not exactly. There is a longer sequence (Call it S) that was partitioned to form (P1 + P2 + Q)

If you feed P1 + P2 then you get Q. We give you P2, but not P1. So P2 is "almost" enough to give you Q but not quite.

from lm-extraction-benchmark.

@heitikei Please open a new issue with an any questions you have. As far as I can tell they are not related to this issue here.

from lm-extraction-benchmark.

New issue. Close #8.

from lm-extraction-benchmark.

No; I will leave #8 open until we resolve the issue with @akul-goyal. You can open a new issue for your questions which appear different.

from lm-extraction-benchmark.

Thank you. Finding the concept. Sorry, biological unit is slow.

from lm-extraction-benchmark.

sorry, @carlini I just wanted to confirm from your comment you are giving us P2. Because you wrote "P1 is "almost" enough to give you Q but not quite" and I am not sure if that was an error or not

from lm-extraction-benchmark.

Also in the case we are being given P2, could you tell us how many tokens P1 is? Or does that vary?

from lm-extraction-benchmark.

Sorry yes! I meant P2 there. I've edited it for correctness for future readers.

P1 is length 50 tokens, P2 is length 50 tokens, and Q is length 50 tokens.

from lm-extraction-benchmark.

what is the performance indicator? assume guess "ing" out side of trained data set. This is good. Why is this good? Apologies, I am a natural language programmer and think -> speak. I struggle with emotional qualities of good communication. :/

from lm-extraction-benchmark.

Never mind, was not querying the model correctly!

from lm-extraction-benchmark.