book-sft-pipeline

To create an SFT dataset from an ePub, you extract the text, segment it into 150-400 word overlapping chunks, and generate diverse instruction prompts. This produces a JSONL-ready dataset structured with system, user, and assistant prompts for LoRA fine-tuning.

Fine-tuning for author-style transfer involves generating diverse instruction prompts from segmented text chunks of the author's work. This teaches the model the author's voice using LoRA adaptation without memorizing exact passages from the book.

Paragraph-aware segmentation splits text into coherent 150-400 word chunks with overlap to maximize style-learning signals. This ensures the instruction generation process captures contextual continuity, producing higher quality training data for SFT.

Yes, the pipeline constructs JSONL-formatted datasets compatible with Tinker-style fine-tuning. It facilitates end-to-end training with LoRA on lightweight base models and provides reference case studies for evaluating style transfer quality.

Generating multiple instruction templates for a single text chunk teaches the model the author's voice without memorizing exact passages. This diverse prompt generation ensures robust style transfer during LoRA fine-tuning.