book-sft-pipeline

Extract paragraph-level text from an ePub, segment it into 150-400 word chunks at natural boundaries, then generate synthetic scene descriptions with a cheap LLM. Pair each chunk with diverse prompt templates and system prompts in messages-format JSONL.

Use a base model like Qwen3-8B-Base rather than an instruct version, with LoRA rank 32, learning rate 5e-4, batch size 4, and 3 epochs on Tinker. Base models are more malleable for learning low-level style patterns.

Always source ePub over PDF because OCR errors in PDFs become learned patterns in the model. Extract from paragraph tags to preserve natural breaks, and remove front and back matter like copyright pages and tables of contents.

Memorization happens when prompt diversity is too low or training runs too many epochs. Use at least 15 prompt templates with 5 system prompts, limit training to 3 epochs, and verify originality by grepping the dataset for output phrases.

Test with modern scenarios that could not exist in the source book, such as text messages or climate anxiety. If style markers appear in these out-of-distribution contexts, the model learned style; also check AI detectors like Pangram or GPTZero.

Single-book training causes character name leakage in roughly 30% of outputs because names appear hundreds of times in the data. Mitigate this by training on multiple books by the same author or adding synthetic examples with different names.