Unfortunately, the templates that I have created are not performing well, the models took this kind of yes/no modeling as a binary classification.
However, I have found the following kind of templates as well:
("What happens next in this paragraph?\n\n{context}\n{options_}", "{answer}"),
("Continue writing the next sentence in this paragraph:\n\n{context}\n\n{options_}", "{answer}"),
("Continue writing the next sentence.\n\n{context}\n\n{options_}", "{answer}"),
("This is a test of commonsense. Complete the next sentence:\n\n{context}\n\n{options_}", "{answer}"),
("Write the next sentence in this paragraph:\n\n{context}\n\n{options_}", "{answer}"),
("How does the next paragraph end?\n\n{context}\n\n{options_}", "{answer}"),
("What most naturally follows?\n\n{context}\n\n{options_}", "{answer}"),
("What happens next?\n\n{context}\n\n{options_}", "{answer}"),
("What is the most logical next event?\n\n{context}\n\n{options_}", "{answer}"),
("Write the next sentence in the following story.\n\n{context}\n\n{options_}", "{answer}"),
these templates play a text generation role so I am thinking of adding this kind of samples during the training process as well!
@jd-coderepos
Unfortunately, the templates that I have created are not performing well, the models took this kind of yes/no modeling as a binary classification.
However, I have found the following kind of templates as well:
these templates play a text generation role so I am thinking of adding this kind of samples during the training process as well!
Any idea about this approach?