How to train my own model
I previously developed a chatbot for intrenal use. The mind is simple: Use IBM watson assistant as the bot and a simple UI. But it is not as good as ChatGPT nowadays. At least it is not easy to tain. The documents have been manually composed to feed it up. So this little project spreaded so many teams.
Then there are a lot of opensource LLM, like llama.cpp, alpaca.cpp, dailai, and gpt4All and I am wondering can I just make a easy-customizing AI tool with these projects. Especially help read documents. Maybe more?
But the first step is how to train or fine-tune a model on my own?
After some intial investigation:
- a A100 GPU or some GPU equavlent required.
- Train data. This would depends on what kind of field you want to go. sentencepiece is a good start
- The tech details of train.
An example exits here
Figure out the tech steps may be the most important step now.