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

Fine-tuning Qwen or a similar instruct model on 192,456 JSONL examples derived from Bitsavers' Microsoft manual corpus (37M+ words, 1977–2005) produces a LoRA adapter that shifts output style toward 1990s technical writing conventions. The pipeline used gemma-4-26b via OpenRouter for paragraph-level quality filtering at roughly $8, with chunks capped at 512 tokens paired with synthetic instructions. The experiment is local-only and the adapter is not being distributed.