I am looking for ways where i create a model that takes as input a block of text(tweets actually), and outputs another block of text that is a modified version of the input. I am looking for ways of doing this, i came across that i could use summarization, but in terms of finetuning it, it probably needs to take in two columns, the text and the summary, but i don’t have the summary, but that’s what I want to generate,
Ideally, its not just summary I want, but a modification to the actual texts, maybe top 5 modifications
The question is , how best can I achieve this
You can use any encoder-decoder (text2text, seq2seq) model for this. Examples are T5, BART.
If you have some training data (input-output pairs), you can use the example scripts to fine-tune the model: transformers/examples/pytorch/translation at main · huggingface/transformers · GitHub (see also the summarization scripts).
Thank you @nielsr for the reply. Yes so for fine-tuning, I know what the inputs are(the input tweets), but as for the outputs, I don’t have them or don’t know what they should be.
as an example, I gave this to chatgpt
modify this tweet “The Ghanaian currency has continued strengthening trend against the US dollar and all the other major foreign currencies. The Cedi in the last two weeks, has been on a path of resurgence against all the major trading currencies. It’s currently trading at ¢8 to a dollar.”
And chatgpt’s response: “The Ghanaian Cedi has seen a sustained trend of appreciation against the US dollar and other major foreign currencies in recent weeks. It is currently trading at an exchange rate of ¢8 to 1 US dollar.”
Something like this is what i want, any suggestions, based on the output and the fine-tuning?