book-sft-pipeline
Use when the user asks to fine-tune on books, create an SFT dataset from books, train a style-transfer or author-voice model, extract ePub text, segment long-form book content, or prepare literary data for LoRA or small-model training.
What this skill does
# Book SFT Pipeline Convert long-form books into supervised fine-tuning data for literary style transfer. Keep the entrypoint lightweight: use this file to route the work, then open only the references needed for the current phase. ## Activate When - Building fine-tuning datasets from literary works - Creating author-voice or style-transfer models - Preparing training data for Tinker or similar SFT platforms - Designing text segmentation pipelines for long-form content - Training small models on limited literary data ## Fast Path 1. Confirm the source is suitable: prefer ePub over PDF, remove front/back matter, and preserve paragraph breaks. 2. Segment into coherent 150-400 word chunks. Never break mid-sentence. 3. Generate synthetic scene descriptions without quoting the source text. 4. Build JSONL conversation examples with varied system prompts and user templates. 5. Train a LoRA on a base model, not an instruction-tuned model. 6. Validate on modern scenarios and grep training data for suspicious phrases. ## Default Parameters | Setting | Default | |---------|---------| | Chunk size | 150-400 words | | Prompt diversity | 15+ templates, 5+ system prompts | | Variants | 2 per chunk | | Model | Qwen/Qwen3-8B-Base or another base 8B-class model | | LoRA rank | 32 | | Epochs | 3 | | Test set | 50 examples minimum | ## Core Rules 1. Source ePub before PDF because OCR noise becomes learned behavior. 2. Keep chunks semantically complete and paragraph-bounded where possible. 3. Teach style, not plot: instructions should describe scenes without quoting. 4. Rotate prompt and system templates to reduce memorization. 5. Use base models for malleable style transfer. 6. Validate originality before claiming the model learned the author's voice. ## Progressive Disclosure Open these only when the task reaches that layer: - [Pipeline Workflow](./references/pipeline-workflow.md) - full phase-by-phase workflow with extraction, segmentation, instruction generation, dataset construction, training, validation commands, costs, and troubleshooting. - [Segmentation Strategies](./references/segmentation-strategies.md) - advanced paragraph, scene, dialogue, and LLM-assisted chunking patterns. - [Tinker Format Specification](./references/tinker-format.md) - Datum, renderer, token-weight, JSONL, and training loop details. - [Tinker API Documentation](./references/tinker.txt) - full API reference. - [Gertrude Stein Case Study](./examples/gertrude-stein/case-study.md) - complete working example with outputs and configuration. ## Implementation Starter Use the sample script when the user wants executable scaffolding: ```bash python plugins/khuym/skills/book-sft-pipeline/scripts/pipeline_example.py ``` The script demonstrates the same pipeline semantics as this skill: segmentation, diverse prompt construction, Tinker datum construction, and originality checks. ## Validation Checklist - Chunks end at natural grammatical boundaries. - JSONL rows contain system, user, and assistant messages. - Prompt variants are distributed across chunks. - Held-out test examples are excluded from training. - Modern scenario outputs contain style markers without original plot content. - Exact output phrases do not appear in the training JSONL. ## References Internal references: - [Pipeline Workflow](./references/pipeline-workflow.md) - [Segmentation Strategies](./references/segmentation-strategies.md) - [Tinker Format Specification](./references/tinker-format.md) - [Tinker API Documentation](./references/tinker.txt) External resources: - [Research Paper](https://arxiv.org/pdf/2510.13939) - [Dataset on Hugging Face](https://huggingface.co/datasets/MuratcanKoylan/gertrude-stein-style-sft) - [Gertrude Stein Case Study](./examples/gertrude-stein/case-study.md) ## Skill Metadata **Created**: 2025-12-26 **Last Updated**: 2025-12-28 **Author**: Muratcan Koylan **Version**: 2.0.0 **Standalone**: Yes
Related in Image & Video
watch
IncludedWatch a video (URL or local path). Downloads with yt-dlp, extracts auto-scaled frames with ffmpeg, pulls the transcript from captions (or Whisper API fallback), and hands the result to Claude so it can answer questions about what's in the video.
physical-ai-defect-image-generation
IncludedUse when the user wants to orchestrate defect image generation, run associated setup, or handle outputs on OSMO. The Day 0 path handles cold-start with USD-to-ROI, image-edit augmentation, and AnomalyGen to create initial PCBA datasets. The Day 1 path performs inference and labeling on real images. This skill helps with first-time asset setup, creation of finetuning checkpoints, and configuring deployment. Trigger keywords: defect image generation, dig workflow, dig pipeline, defect image detection workflow, aoi pipeline, aoi anomalygen, usd2roi anomalygen, day 0 pcba, day 1 pcba, day 1 real-photo alignment, day 1 manual roi, metal surface anomaly, glass defect, anomalygen finetune, setup_pcb, setup_metal, setup_glass, setup_pretrained, dig setup, dig datasets, dig pretrained checkpoint, dig image-edit endpoint.
accelint-react-best-practices
IncludedReact performance optimization and best practices. ALWAYS use this skill when working with any React code - writing components, hooks, JSX; refactoring; optimizing re-renders, memoization, state management; reviewing for performance; fixing hydration mismatches; debugging infinite re-renders, stale closures, input focus loss, animations restarting; preventing remounting; implementing transitions, lazy initialization, effect dependencies. Even simple React tasks benefit from these patterns. Covers React 19+ (useEffectEvent, Activity, ref props). Triggers - useEffect, useState, useMemo, useCallback, memo, inline components, nested components, components inside components, re-render, performance, hydration, SSR, Next.js, useDeferredValue, combined hooks.
elevenlabs-agents
IncludedBuild conversational AI voice agents with ElevenLabs Platform using React, JavaScript, React Native, or Swift SDKs. Configure agents, tools (client/server/MCP), RAG knowledge bases, multi-voice, and Scribe real-time STT. Use when: building voice chat interfaces, implementing AI phone agents with Twilio, configuring agent workflows or tools, adding RAG knowledge bases, testing with CLI "agents as code", or troubleshooting deprecated @11labs packages, Android audio cutoff, CSP violations, dynamic variables, or WebRTC config. Keywords: ElevenLabs Agents, ElevenLabs voice agents, AI voice agents, conversational AI, @elevenlabs/react, @elevenlabs/client, @elevenlabs/react-native, @elevenlabs/elevenlabs-js, @elevenlabs/agents-cli, elevenlabs SDK, voice AI, TTS, text-to-speech, ASR, speech recognition, turn-taking model, WebRTC voice, WebSocket voice, ElevenLabs conversation, agent system prompt, agent tools, agent knowledge base, RAG voice agents, multi-voice agents, pronunciation dictionary, voice speed control, elevenlabs scribe, @11labs deprecated, Android audio cutoff, CSP violation elevenlabs, dynamic variables elevenlabs, case-sensitive tool names, webhook authentication
humanizer
IncludedHumanize AI-generated text by detecting and removing patterns typical of LLM output. Rewrites text to sound natural, specific, and human. Uses 28 pattern detectors, 560+ AI vocabulary terms across 3 tiers, and statistical analysis (burstiness, type-token ratio, readability) for comprehensive detection. Use when asked to humanize text, de-AI writing, make content sound more natural/human, review writing for AI patterns, score text for AI detection, or improve AI-generated drafts. Covers content, language, style, communication, and filler categories.
generating-mermaid-diagrams
IncludedSalesforce architecture diagrams using Mermaid with ASCII fallback. Use this skill when generating text-based diagrams for Salesforce architecture, OAuth flows, ERDs, integration sequences, or Agentforce structure. TRIGGER when: user says "diagram", "visualize", "ERD", or asks for sequence diagrams, flowcharts, class diagrams, or architecture visualizations in Mermaid. DO NOT TRIGGER when: user wants PNG/SVG image output (use generating-visual-diagrams), or asks about non-Salesforce systems.