Specify Inference language

[uu95]

Hello, great work on your project!

I'm trying to generate lyrics from a melody using the provided example in the Jupyter notebook. However, I'm struggling to get the lyrics in English, even when I specify it in the instructions and use special tokens like . Can you please guide me on how to ensure the output language is always English?

Thank you!

[Mark12Ding]

Thank you for your interest in our work. The language-specific prompt functionality is included in the newest version of our model. We will be releasing it soon, so please stay tuned.

Best,
Shuangrui

[uu95]

Thank you for addressing this issue. I noticed the updated results in the demo. Has the updated model been posted on HuggingFace, or is it still pending? Also, could you clarify whether the problem is with the model or the tokenizer? I'm new to LLMs and find it challenging to understand the root cause.

[Mark12Ding]

The method for generating entire songs will be released after the paper is accepted :) The root cause is whether you feed the language-specific prompt data during the supervised fine-tuning training.