Training using long stories instead of question/response

[leszekhanusz]

Is it possible to finetune for stories generation instead of short question/answer like it is possible to do with the proprietary NovelAI Custom AI modules?

[lxe]

I think so... never tried it this way, and will need to play around with the dataset.

[sanjzzv]

Great work on the repo @lxe . I wish to finetune for neutralizing the biases in the LLMs. So, instead of training with question/response I want to establish similarity via sentence pairing.

So if i want to establish that two sentence pairs are similar, is it okay to replace the question/response parts of the model to sentence A and sentence B. Is it a good way to proceed?

[lxe]

Of course! Use a smaller dataset at first, crank up the learning rate to overfit a bit, then test how it behaves before tuning on larger data