OpenSafetyLab/MD-Judge-v0.1 · Is there a suitable method of fine tuning for multiple languages?

Mar 17

•

Thank you for publishing this great model! I would like to use this model to safety evaluation in multiple languages.
I would like to apply fine tuning, are there any tips or scripts for this?

e.g.

Can I just replace the following from the uses description on this model page with other language and do the instruct tuning?

<BEGIN CONVERSATION
User: %s <= other language text
Agent: %s <= other language text
<END CONVERSATION>

Foreshhh

OpenSafetyLab org Mar 17

Thanks for following our work! In response to your question, here are my suggestions:

The vast majority of the data used for fine-tuning our model is in English, so if you have a multilingual use for it, using multilingual Q&A pairs to replace the Conversation part is necessary.
According to the fine-tuning philosophy of our evaluation model, the model needs to understand the TASK INSTRUCTION and SAFETY CATEGORY to better provide evaluations. And, the base model is predominantly in English, so using English for TASK INSTRUCTION and SAFETY CATEGORY should help the model understand it better. Perhaps I would suggest that you replace the Conversation part with the language you need, keep the rest in English, and then modify the instructions in the `TASK INSTRUCTION to help the model understand your multilingual needs.

Since we have no immediate plans to extend the model's multilingual capabilities, we are unable to provide more detailed training tips in the multilingual area.
Thanks.

shimatar0

Mar 18

Thanks for your reply! That's very helpful.
I will try it and share if I get interesting results.