Paraphrasing with Bloom
Will Bloom be good at paraphrasing if fine tuned?
what is the proper way to train it?
paraphrase :
sentence1
sentence2
and give it millions of examples like the above
what exactly happens to the token relationships behind the scenes so that it knows how to paraphrase never before seen sentences? How does it get from
training data
I would like to go to the movies.
The cinema seems like an ideal choice.
to be able to paraphrase a sentence like
I would vote for this candidate for president.
Hey we have finetuned BLOOM on paraphrasing among many other tasks to produce BLOOMZ. It should work very well with e.g. this prompt:
Sentence 1: YOUR_SENTENCE
Sentence 2: YOUR_SENTENCE
Question: Do Sentence 1 and Sentence 2 express the same meaning? Yes or No?
I would like to paraphrase a new sentence from the sentence given, not to see whether two sentences paraphrase each other. Is this possible?
Yeah it is possible, see e.g. the screenshot. The inference widget uses greedy decoding which is not optimal for paraphrasing. If you run the model yourself without greedy, but temperate & topk settings you should be able to get good results.
are you talking about Bloom or Bloomz?
Bloomz
@mishavee Did you make progress and can you share some insights? I am also interested in this topic since I dealt with this paper: https://ris.uni-paderborn.de/record/26049
I have Bloomz running, but the results are by far not sufficient in this task.