book-sft-pipeline

Extract text from an ePub, segment it into 150-400 word chunks at paragraph boundaries, generate diverse synthetic instructions for each chunk, and train a base model with LoRA on Tinker. Use at least 15 prompt templates to prevent memorization of specific phrasings.

Segment the book into coherent chunks, use a cheap LLM to generate scene descriptions without quoting the text, then pair each description with the original passage in messages format. Generate two variants per chunk with different system prompts and templates.

Use base models like Qwen3-8B-Base rather than instruction-tuned versions. Base models are more malleable for learning new styles, while instruct models have existing patterns that resist being overwritten by style training.

Exact phrase repetition indicates too few prompt variations or too many training epochs. Use 15 or more prompt templates with 5 or more system prompts, and limit training to 3 epochs to force the model to learn style patterns instead of memorizing content.

Chunks of 150-400 words with paragraph overlap produce more training examples and better style transfer than larger 250-650 word chunks. Always break at paragraph boundaries, never mid-sentence, to avoid teaching fragmented output.

Approximately $2 total: about $0.50 for LLM-based instruction generation and $1.50 for roughly 15 minutes of Tinker LoRA training. This yields 500-1000 training examples from a typical novel.