What problem does it solve?
This Skill turns long-form books into high-quality supervised fine-tuning (SFT) datasets for style transfer, so models learn author voice instead of memorizing plot.
Core Features & Use Cases
- Book → training text extraction: Extract clean paragraph-level text from ePub (favoring ePub over PDF to avoid OCR artifacts) and remove front/back matter.
- Intelligent segmentation for training: Split into semantically coherent chunks (typically 150–400 words) with overlap and safeguards to avoid mid-sentence or poorly formed breaks.
- Diverse instruction generation: Create many instruction/prompt variations (15+ templates and multiple system prompts) so the model learns style patterns rather than a single prompt mapping.
- Dataset construction for Tinker-style training: Build chat-message JSONL pairs (system/user/assistant) suitable for supervised tuning.
- LoRA SFT training + validation: Train a small base model with LoRA, then validate using modern scenario tests and originality checks.
- Use case examples: Fine-tune on books to replicate an author’s voice, build LoRA adapters for style writing, and prepare Tinker-compatible datasets from long novels.
Quick Start
Activate this skill when the user asks to fine-tune a model on a specific book’s author voice, and then follow the pipeline from ePub text extraction through segmentation, instruction generation, JSONL dataset building, LoRA training, and modern-scenario originality validation.
