Book Training: Style Transfer via SFT

A complete pipeline for training language models to write in a specific author's style using Supervised Fine-Tuning (SFT) with LoRA.

Project Overview

This repository demonstrates how to:

1. Extract and segment text from EPUB books
2. Generate diverse instruction-response pairs using LLMs
3. Train a small model (Qwen3-8B-Base) with LoRA
4. Validate genuine style transfer vs. memorization

Results

Trained on Gertrude Stein's "Three Lives" (1909):

Metric	Value
Training Examples	591
Final Test Loss	213 (from 7,584)
Loss Reduction	97%
Training Time	~15 min
AI Detector Score	~50-70% Human (Pangram)

The model applies Stein's style to modern scenarios not in training data, confirming style transfer rather than memorization.

Repository Structure

├── skills/book-sft-pipeline/     # Reusable skill for AI agents
│   ├── SKILL.md                  # Main skill definition
│   ├── references/               # Segmentation, training docs
│   ├── scripts/                  # Example pipeline code
│   └── examples/gertrude-stein/  # Complete case study
├── src/
│   ├── pipeline.ts               # Dataset generation pipeline
│   └── train-tinker.ts           # Tinker training setup
├── train_tinker.py               # LoRA training script
├── chat.py                       # Interactive model testing
├── test_style_transfer.py        # Style validation tests
├── hf-dataset/                   # HuggingFace dataset files
│   ├── train.jsonl               # 591 training examples
│   ├── test.jsonl                # 49 test examples
│   └── README.md                 # Dataset card
└── three-lives-sft-dataset.jsonl # Full generated dataset

Quick Start

1. Generate Dataset

npm install
npx tsx src/pipeline.ts

2. Train with Tinker

pip install tinker python-dotenv
python train_tinker.py

3. Test the Model

python chat.py

Dataset

Available on HuggingFace: MuratcanKoylan/gertrude-stein-style-sft

from datasets import load_dataset
dataset = load_dataset("MuratcanKoylan/gertrude-stein-style-sft")

Key Techniques

Prompt Diversity

15 prompt templates × 5 system prompts = 75 unique combinations per chunk. Prevents overfitting to specific phrasings.

Fine-Grained Segmentation

150-400 word chunks with overlap. Smaller chunks capture more stylistic patterns.

Modern Scenario Testing

Validate style transfer by prompting about topics the author never wrote about (real estate offices, smartphones, etc.).

Skills for AI Agents

The /skills/book-sft-pipeline folder contains a complete skill definition that enables AI agents to replicate this pipeline for any book:

skills/book-sft-pipeline/
├── SKILL.md              # Agent-readable instructions
├── references/           # Technical documentation
├── scripts/              # Example implementations
└── examples/             # Case studies with results

Sample Output

Prompt: Write about a tech startup founder in the style of Gertrude Stein.

Output:

She was always working, always she was working on her startup. The investors they would come and they would go and she would talk to them about the product. She was a good founder, always she was a good founder. She had this way of explaining things, simple and direct, and the investors they would listen.

License

MIT

Source text "Three Lives" by Gertrude Stein (1909) is in the public domain.