About fine-tuning

[fukuzawa-e]

Are there codes for fine-tuning a new model by my own training data ?
Is it possible to use training data in Japanese (text data) directly ?

[ImKeTT]

Thanks for your interest in our work and sorry for the late reply. The released code does not include fine-tuning the LM for generation. But technically, you can use any fine-tuned LMs for ZeroGen, including Japanese LMs.
But also note that, if you want to use Japanese LM for ZeroGen, you should also change the CLIP and GloVe model to the Japanese version.

[fukuzawa-e]

I see. Thank you so much for your guidance.

[fukuzawa-e]

I have another question about fine tune.
I found there is a language mode (cambridgeltl/magic_mscoco), if I plan to use Japanese model, should I re-train a Japanese language model based on cambridgeltl/magic_mscoco ?
Is there any training source on how to training a new language mode for Japanese ?

[ImKeTT]

For the first question, the answer is no, you don't need to train a Japanese LM based on cambridgeltl/magic_mscoco. It's enough to just fine-tune a pre-trained Japanese LM on the textual captions of any caption datasets you want.

For the second question, I believe there're plenty of open-source codes on GitHub about fine-tuning Japanese LMs. Unfortunately, I'm not an expert of Japanese or Japanese LM, so I have no experience on fine-tuning a Japanese LM. I'm afraid you need to find the code and model by your own then.

But please don't hesitate to ask any questions that you may encounter when applying ZeroGen to other LMs, and you're always welcome to open a PR to integrate them into ZeroGen :)

[fukuzawa-e]

I see. Thank you so much for your information.