Training on own data?

[dinnerisserved]

Is there a way to feed GPT4all own data so that it can be trained on the information? I would like to be able to feed it my emails, my PDF files and a bunch of other data that I have, and use GPT4all's chat to trawl through this data and spit out information for me. Is this something that is being worked on, or currently possible?

[cigoic]

vote this! Same question. ✋

[mdogadailo]

LangChain allows you to connect a language model to other sources of data. you can also take a look https://github.com/imartinez/privateGPT

[mehrooo]

Its known as "Chat with PDF" or "Talk to PDF/Book/Document", look it up
You don't need to train a model on your own data, you can finetune a model in a way that it analyse the given file and extract the asked information from it.

[dinnerisserved]

I never intended to "train" on own data, but it was more about letting the GPT access a file repository to take into consideration when asking it questions. Chat with a datalake is what I wanted to achieve.

[sunilkumardash9]

I never intended to "train" on own data, but it was more about letting the GPT access a file repository to take into consideration when asking it questions. Chat with a datalake is what I wanted to achieve.

Well, I think you can do this by performing a semantic search over your text data (embeddings) and feed the relevant ones to chat models and get your answers.

[naman-TI]

Semantic search is one option agreed. But it sounds better than how it actually performs.

Is there a way to fine-tune (domain adaptation) the gpt4all model using my local enterprise data, such that gpt4all "knows" about the local data as it does the open data (from wikipedia etc)

[zubair-ahmed-ai]

I never intended to "train" on own data, but it was more about letting the GPT access a file repository to take into consideration when asking it questions. Chat with a datalake is what I wanted to achieve.

Well, I think you can do this by performing a semantic search over your text data (embeddings) and feed the relevant ones to chat models and get your answers.

Tried that with dolly-v2-3b, langchain and FAISS but boy is that slow, takes too long to load embeddings over 4gb of 30 pdf files of less than 1 mb each then CUDA out of memory issues on 7b and 12b models running on Azure STANDARD_NC6 instance with single Nvidia K80 GPU, tokens keep repeating on 3b model with chaining
Currently trying to use the mosaicml/mpt-7b to generate some tokens and that's taken forever.

Came here to test out GPT4All to see if this is any better

[claell]

@niansa: #87, #198, #223

[niansa]

yes, this is a duplicate.

please open a new, updated issue if this is still relevant to you.
You're encouraged to open new issues or even PRs if there's anything you need!