How to train and adapt to a new language and keep editing ability of the base model ?

[xiaoyangnihao]

How to train and adapt to a new language and keep editing ability of the base model, by the way, how many hours of new language is at least needed.

[hldwby]

For new language tasks, you can use SFT and add language tags at the beginning of sentences in the training data. In our experience, 100 hours of data can achieve basic performance, while 1,000 hours can reach usable performance.
To retain the editing capability, we suggest including some editing data in the SFT process.

[xiaoyangnihao]

For new language tasks, you can use SFT and add language tags at the beginning of sentences in the training data. In our experience, 100 hours of data can achieve basic performance, while 1,000 hours can reach usable performance. To retain the editing capability, we suggest including some editing data in the SFT process.
@hldwby any plan to release a small editing datasets to keep the editing ability when finetuning?

[hldwby]

For new language tasks, you can use SFT and add language tags at the beginning of sentences in the training data. In our experience, 100 hours of data can achieve basic performance, while 1,000 hours can reach usable performance. To retain the editing capability, we suggest including some editing data in the SFT process.
@hldwby any plan to release a small editing datasets to keep the editing ability when finetuning?

Sorry, we currently have no plans to release a small editing dataset.
However, you can use the editing capabilities of Step-Audio-EditX to generate a set of paired data and incorporate it into your SFT pipeline. Just make sure to check the quality of the paired data before training.
BTW, we also use synthetic paired data for our model training, and this approach works well.