The developers used the new open source Meta language model “LLaMA 7B”, which has not yet reached the full version. It is the smallest and cheapest technology available today.

What is known about alpaca

One of the versions of GPT, the core underlying ChatGPT, was used to train Alpaca. Requested 175 human-written instruction/result pairs, and proceeded to produce 20 more of them in the same style and format, each time. This was automated using one of OpenAI’s open APIs. Thus, in a short time the team had about 52,000 speech samples that could be used to train the new LLaMA model.. The creation of this huge training data, 500.

All this data was then used to fine-tune the LLaMA model. The process took approximately three hours on eight A100 cloud processing computers with 80 gigabytes of capacity. less costly 100 USD.

The developers warn that they have not configured Alpaca to be safe and harmless, so they ask anyone who installs it to report any security or ethical issues. However, it is already clear that this neural network has a lot of potential. The first code of the project has already been unlocked, thanks to the enthusiasts being able to work on the development. One of the most remarkable features of Alpaca is that it can work in offline mode without Internet access, and this neural network can work on any computer, including single-board computers, which does not differ in high performance.

The authors claim that everything can be done even cheaper, which means: an unlimited number of unsupervised language patterns will now appear.