book-sft-pipeline

Convert books into SFT datasets and orchestrate LoRA training pipelines.

Updated May 24, 2026

Search Tags:#dataset #epub #book #style-transfer #SFT #loRA

Ready to run

npx skills add https://github.com/FVossebeld/agent-skills-for-context-engineering --skill book-sft-pipeline-fvossebeld

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: book-sft-pipeline
Source: https://github.com/FVossebeld/agent-skills-for-context-engineering/tree/main/examples/book-sft-pipeline
Command: npx skills add https://github.com/FVossebeld/agent-skills-for-context-engineering --skill book-sft-pipeline-fvossebeld

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) components.

What problem does it solve?

This skill automates the end-to-end process of converting books into training-ready SFT datasets and running style-transfer LoRA pipelines, reducing manual setup and repetitive work.

Core Features & Use Cases

ePub extraction and text normalization: convert books into clean text suitable for segmentation.
Intelligent segmentation: split text into 150-400 word chunks with deterministic overlap to maximize style signal.
Instruction generation with prompts templates: generate diverse instructions and prompts to encourage robust style transfer.
Dataset construction: assemble JSONL datasets compatible with Tinker for supervised fine-tuning, including system/user/assistant messages.
LoRA training workflow integration: orchestrate base-model fine-tuning with 32-rank adapters on common base models, plus evaluation hooks.

Quick Start

Provide an end-to-end example pipeline to train a Gertrude Stein style on a given book using the included SKILL.md and references.

Frequently Asked Questions about book-sft-pipeline

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I create an SFT dataset from an ePub book for style transfer?▼

SFT datasets for style transfer require segmenting book text into 150-400 word chunks with deterministic overlap, generating diverse instructions, and assembling JSONL files with system, user, and assistant messages for supervised fine-tuning.

Can I use LoRA adapters to train an author style transfer model on Qwen3?▼

You can train an author style transfer model using LoRA adapters on base models like Qwen/Qwen3-8B-Base, orchestrating fine-tuning with 32-rank adapters and built-in evaluation hooks.

What is the best way to segment book text for training a style transfer model?▼

The best way to segment book text for style transfer models is splitting it into 150-400 word chunks with deterministic overlap to maximize the captured style signal.

Does this dataset construction pipeline support JSONL output for Tinker?▼

Yes, the dataset construction pipeline supports JSONL output for Tinker by assembling supervised fine-tuning data with structured system, user, and assistant messages.

What are the limitations of using fixed chunk sizes for ePub text segmentation?▼

Fixed chunk sizes of 150-400 words for ePub text segmentation may not perfectly align with narrative boundaries, though deterministic overlap is applied to preserve context and maximize style signal.