IMHO the `README.md` for instructlab itself does a terrible job of explaining the value of the software. I found the taxonomy `README.md` much more helpful: https://github.com/instructlab/taxonomy/blob/main/README.md

One of the authors here. This is the paper backing up the technology.

https://arxiv.org/abs/2403.01081

This seems like a really nice framework for feeding documentation into llms.

Would be curious to use same dataset without fine-tuned llm for RAG over same data.

Then one could immediately make use of building the dataset and then measure gains from training.