Can It work on Multilingual data?

[Komal-99]

Hi,
I want to ask that PrivateGPT only works on English data or can work on French, German dataset?
For ex- We have a corpus of French language, and we made its embeddings using Ingest.py file then my model will be able to create responses in French?

Model using - Llama--7B-chat

[Lue-C]

Hi,
I am very interested in building a version of privateGPT that processes non-english texts and gives non-english answers (in my case it would be German). Processing non-english data requires an embedding model, that was trained on the desired language. I tried "distiluse-base-multilingual-cased-v1" and it worked ok (but much better than the default embedding model). Yet there is much room for improvement on the choice.

For getting non-english answers a llm is required which was trained on the desired language. So far I did not find a good overview which models are supported by GPT4All and LLama. Do you know one?
At the moment I am using the "GPT4All-13B-snoozy.ggmlv3.q4_0" model and found that for some prompts it does give German answers and for others it does not. This could be adressed by changing the default prompt.

Did you find a solution in the meantime?

[Komal-99]

Actually, PrivateGPT is just a method to use the LLM model, so It depends on which model you are using, I have overcome those language Issues for German by using vicuna 13b ggml model. It works quite well on both English and German text.

[Lue-C]

I searched for it but could not find a bin file. Where can I download it?

[Komal-99]

https://huggingface.co/TheBloke/Vicuna-13B-v1.3-German-GGML

[Lue-C]

Thanks a lot, this works really good:)

[Komal-99]

Most welcome , If you cane across any multilingual model please so let me know which can support spanish German french english all.

[NadaComa]

Hi,
so can you confirm that using distiluse-base-multilingual-cased-v1 for embedding and vicuna as LLM gives satisfying results for german documents?

[NadaComa]

Yes, I do. Interested in sharing experiences in a call?

[Lue-C]

Yes, that is a great idea. Which app do you prefer?

[NadaComa]

Whatsapp or Telegramm or Signal

[Lue-C]

Then let us meet via telegram, you can send me your username to [email protected]
When do you have time this week?

[NadaComa]

Did you receive my mail?

[Komal-99]

@Lue-C which all languages you have tested with distiluse-base-multilingual-cased-v1 ?
Did you found any another for multilingual as vicuna is only for German

[Lue-C]

Not yet, but I have not been looking for other models in the last week because of other priorities. I will write if there is progress on that matter.
A possible way around working with several languages could be to detect the language of the prompt and then use a language specific model. Do you think this is feasible?

[Komal-99]

Yes. I am also looking over this approach