Fine-tuning an LLM to write docs like it's 1995

The author fine-tuned a language model to write documentation in the style of 1990s technical writers by training it on over 37 million words of scanned Microsoft manuals from Bitsavers, a repository of vintage computer documentation. After cleaning and preparing the data into 192,456 training examples, they used fine-tuning rather than full training to achieve the style transfer, demonstrating that specialized, locally-run LLMs for specific writing styles are feasible alternatives to relying on more powerful cloud-based models.