Book Style SFT Pipeline
Turn eBooks into author-style SFT datasets, then train and evaluate LoRA models for style transfer.
The documentation describes staged artifacts, validation steps, and some failure symptoms, but it does not define user confirmation, permission boundaries, data flows, copyright or sensitive-text handling, external-service data use, credential management, or rollback. It recommends sending book text to an external LLM and training through Tinker, so substantial points are deducted. MIT licensing and author metadata provide limited attribution, while publisher identity remains unverified.
The main phases, sample code, configuration, and known issues make the happy path plausible. However, dependencies are not pinned or fully installed, the Tinker renderer/model examples may be mismatched, segmentation and overlap logic can exceed target bounds, and abnormal-input handling and recovery are thin. This is a static review with no committed tests covering the selected skill's key paths; under calibration, reliability cannot exceed 10, so points are deducted.
Activation scenarios, intended audience, and main inputs and outputs are reasonably clear across ePub extraction, segmentation, dataset construction, LoRA training, and evaluation. Non-fit boundaries, copyright constraints, alternatives, and exclusion triggers are under-specified. Core use depends on Gemini, Tinker, and overseas model services, with no mainland-China reachability guidance, so environment-fit uncertainty reduces the score.
SKILL.md, README, references, scripts, the case study, and version metadata provide useful progressive documentation with installation guidance, configuration, known limitations, and an MIT license. Full marks are not justified because maintenance ownership, update path, changelog, pinned dependencies, and systematic FAQs are missing, and some documented results are not fully consistent with the code/configuration.
The skill covers the conceptual book-to-SFT workflow and supplies adaptable code, configuration, and a case study. However, the script is explicitly conceptual, the case study reports only about 50% perfect transfer and about 30% name leakage, and results depend on external platforms and manual adaptation. The output is therefore not directly usable without substantial review. Static calibration caps effectiveness at 7, so points are deducted.
Configuration, loss tables, sample outputs, validation commands, and limitation disclosures provide some auditability. Nevertheless, the evidence is primarily author-reported repository material, without independently reproducible execution records, committed tests for the selected skill's key paths, or meaningful cross-source corroboration. Static calibration caps verifiability at 5, so only limited credit is awarded.
- Before training on copyrighted books or replicating an author's style, confirm authorization, applicable law, and platform policy; the documentation does not define these boundaries.
- Gemini, Tinker, model downloads, and AI-detector services may expose training text or credentials and may have reachability limitations; review data flows, permissions, terms, and mainland-China availability first.
- The reported results are primarily author-provided and include roughly 50% perfect transfer, 30% character-name leakage, and 5% failure; independent reproduction and regression testing are required before deployment.
- The code and configuration carry platform/renderer compatibility, segmentation-boundary, dependency-version, and error-handling risks and should not be used in production training without adaptation.
What does this skill do, and when should you use it?
This skill documents an end-to-end workflow for extracting books from ePub, segmenting literary text, generating synthetic instructions, and building JSONL training data. It then covers LoRA training on Tinker, using Qwen/Qwen3-8B-Base with rank 32 and three epochs as the example configuration. The approach emphasizes natural 150–400-word segments, diverse prompts, and modern-scenario tests that distinguish style learning from content memorization. It also addresses character-name leakage, phrase parroting, and fragmented outputs.
Reads chapters and paragraphs from ePub files, removes front and back matter, and produces plain text; segments text at paragraph boundaries into overlapping 150–400-word chunks; calls a fast LLM to generate scene descriptions without directly quoting the excerpt; builds message-format JSONL examples with multiple system prompts and prompt templates; runs LoRA training on Tinker with Qwen/Qwen3-8B-Base; and validates outputs with modern scenarios, training-data phrase searches, and AI detectors.
- A researcher building an author-voice SFT dataset from a literary eBook can use the extraction, segmentation, and instruction-generation workflow.
- A team training an 8B-or-smaller model for literary style transfer can use the documented LoRA configuration and training loop.
- A developer trying to prevent plot memorization can apply the diverse-prompt strategy and modern-scenario evaluation.
- An engineer generating JSONL training data from long-form fiction can follow the staged, resumable pipeline design.
What are this skill's strengths and limitations?
- Covers the workflow from ePub text to training and validation.
- Includes concrete Python and Tinker-style examples.
- Gives specific guidance on paragraph boundaries, 150–400-word chunks, and prompt diversity.
- Includes modern-scenario testing, originality checks, and failure-mode remedies.
- The workflow depends on ePub files, Python example libraries, Tinker, an external LLM, and the specified base model; the source does not provide a complete environment setup.
- The example focuses on one book and author style, so character leakage and phrase reproduction can still occur.
- The source provides no independent test suite or cross-platform test evidence.
- The stated cost, training time, and style-transfer rate are documented expectations, not independently verified results.
How do you install this skill?
The README states that skills should be installed as directories rather than flattened into a single SKILL.md file. Clone the repository with git clone https://github.com/muratcankoylan/Agent-Skills-for-Context-Engineering.git, then copy examples/book-sft-pipeline/ into the target Agent Skills directory. The source does not provide a dedicated installation command or complete dependency setup for this example.
How do you use this skill?
Load the directory in an Agent Skills-compatible client and activate it when building literary-work SFT data, creating author-style models, designing long-text segmentation, or evaluating style transfer. Follow the extraction, segmentation, instruction-generation, dataset-construction, LoRA-training, and validation phases. The source does not specify a client-specific trigger syntax.
How does this skill compare with similar options?
The source compares its 150–400-word overlapping segmentation approach with a larger 250–650-word method and claims it produces more examples; it does not name a competing complete product or framework.