💡 [REQUEST] - Is it possible to train the model with a new language

[rajat-ventura]

起始日期 | Start Date

No response

实现PR | Implementation PR

No response

相关Issues | Reference Issues

No response

摘要 | Summary

Thanks for the great work

I am looking for a model that can do automatic audio detection in realtime for indic languages. Is it possible for me to train this model on new Indian regional languages like Hindi, Marathi, Tamil, Telugu for starters. I have some raw dataset with me to try. If possible can you please guide me with the relevant steps?

基本示例 | Basic Example

For example,

if I say "कैसे हो" it should be able to transcribe it in hindi in the same manner

缺陷 | Drawbacks

One drawback maybe is that the pretrained language dialects might interfere in between while doing the live detection

未解决问题 | Unresolved questions

No response

[tc-mb]

Hello, in principle, it's definitely possible. However, it depends on the availability of data. With a rich corpus of a new language as a reference, training to achieve good results is feasible.

Regarding training with a new language, I suggest you first fine-tune the model using a small amount of corpus to see its effect. If the effect is significant, then training with a dataset of about the size of a SFT might be sufficient. However, if the effect is not significant, it's possible that this version of LLM has a slightly weaker understanding of the language you need, in which case the training workload will be on par with the pre-training workload, much larger. I recommend conducting a small-scale experiment first to confirm. For a fine-tuning tutorial, please refer to our Cookbook: https://github.com/OpenSQZ/MiniCPM-V-CookBook