So assuming I download some models*.bin (like BERT, llama-2) on my CPU platform locally after training on GPU on Colab.
What I do not understand is following: The model was pretrained on xy datasets and is able to provide coding, nlp output, as tested on colab.
After that will be fine-tuned on the own datasets.
How will the model from this point work? Meaning does it need access to the datasets that it was trained to provide the expected coplex multi-tasking answers?
What i imagined is to use the pre-training as a base "understanding" and train it on own datasets.
Then provide access to the own Datasets. These should have a use of a basic memory or knowledge base of the model (How to think). Example dataset colection 1 : reasoning, Q/A, Indexes, Lexicon dataset colection 2 : own data (files, documentations, manuals, projects...)
Would a database fit better for this purpose?
What did I misunderstand on the whole concept?