book-sft-pipeline

The book-sft-pipeline skill provides a comprehensive architecture for developers and writers aiming to fine-tune small language models (8B or less) to emulate specific authorial voices. It focuses on high-fidelity data processing, ensuring that models learn the rhythm, vocabulary, and stylistic nuances of a literary source rather than just memorizing plot points. This skill is ideal for projects involving digital humanities, creative writing assistants, or voice-replication models where text-based style transfer is the primary objective.

• Automated ePub extraction using BeautifulSoup to parse paragraph structures while removing front and back matter that pollutes training datasets.

• Intelligent text segmentation that enforces natural boundaries, targeting 150-400 word chunks with overlapping sequences to maintain semantic coherence.

• Diverse synthetic instruction generation using multiple system prompts and varied templates to prevent overfitting and encourage generalizable style acquisition.

• Optimized data preparation for Tinker and standard SFT platforms, outputting structured JSONL files suitable for LoRA (Low-Rank Adaptation) training.

• Integrated validation methodologies, including AI detection and originality checks to ensure the generated synthetic data adheres to the intended tone.

• Always prioritize ePub source material over PDF to avoid common OCR-related errors that manifest as hallucinations during training.

• Utilize the orchestrator agent pattern to manage the state across the four-phase pipeline: extraction, segmentation, instruction synthesis, and dataset building.

• Apply the 15+ prompt template and 5+ system prompt strategy to ensure the model learns stylistic patterns across varying contexts.

• Target small-scale deployments; this pipeline is specifically calibrated for efficient LoRA training on limited data samples (approx. 500-600 examples).

• Be mindful of context engineering principles: focus on high-signal data curation rather than raw volume, adhering to attention mechanics that favor quality over quantity to mitigate context degradation.