Fine-tuning an LLM to write docs like it's 1995
4 hours ago
- #QLoRA
- #tech writing style
- #LLM fine-tuning
- Author fine-tuned an LLM to write technical documentation in a 1990s Microsoft style.
- Used Bitsavers' scanned manuals as a 37-million-word corpus, cleaned via OCR and classification.
- Fine-tuned with QLoRA on Runpod, testing Llama 3.1 8B and Qwen 2.5 7B models under varied conditions.
- Fine-tuned models successfully adopted period-specific structures and vocabulary in tests.
- Experiments showed style transfer works, but models lack judgment and require human oversight.