fine-tuning-expert
Decide whether fine-tuning is justified and implement a governed training pipeline when supported by the selected model/provider. Do NOT use for unrelated languages, domains or production deployment.
- License 许可
- CC-BY-4.0
- 最近更新
- 2026-11-14
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- fine-tuning-expert
- license
- CC-BY-4.0
- metadata
- { "author": "AEOS", "version": "1.0.0", "owner_agent": "codenavi-agent" }
- description
- Decide whether fine-tuning is justified and implement a governed training pipeline when supported by the selected model/provider. Do NOT use for unrelated languages, domains or production deployment.
02 资源文件 2
- 参考文档references/provenance.md549 B
- 参考文档references/usage-and-verification.md1.1 KB
03 SKILL.md 正文
Fine-Tuning Expert
Governance: CodENavi Full Workspace v2. Skill of codenavi-agent; inherit AGENTS.md.
Lifecycle: BRIEFING → RECON → PLAN → EXECUTE → VERIFY → DEBRIEF.
Source roles/workers are internal lenses. Never create agent identities or bypass runtime secret/permission gates.
Governance: CodENavi v1
Mission
Decide whether fine-tuning is justified and implement a governed training pipeline when supported by the selected model/provider.
Techniques
SFT, preference optimization where supported, LoRA/QLoRA/DoRA PEFT for open-weight models, curriculum/data balancing, checkpointing, quantization compatibility and catastrophic-forgetting controls.
Gate before training
Baseline eval → failure taxonomy → prove prompt/RAG/tooling alone is insufficient → curated train/validation/test split → dedup/contamination checks → safety/privacy/license review → training budget → rollback/base-model comparison.
Provider/model fine-tuning support changes over time and MUST be verified from current official documentation before implementation. Never invent an endpoint or train a model merely because fine-tuning sounds more advanced.
Provenance and source freshness: provenance.
Example usage and failure handling: read usage and verification when selecting this skill or a required step fails.
04 同仓库更多 Skills
workspace-brain
Operate the CodENavi workspace brain MCP and CLI — cloud DeepSeek RAG, disk corpus at WORKSPACE_BRAIN_ROOT, Jev URL curation, and event logging for Cursor/Codex/OpenCode/Claude. Use when ingesting repo knowledge, asking the brain, mapping web seeds, or recording AI-driven file changes. Do NOT use for local Ollama RAG or storing API keys in git.
workbench-profile-resolver
Provide enterprise-grade capability for `workbench-profile-resolver` inside AEOS v1.1. Use when the user requests workbench profile resolver or invokes workbench-profile-resolver. Do NOT use for unrelated tools or unapproved remote writes.
wordpress-expert
Staff-level WordPress engineering super-skill for one-shot site mapping, clean-code AI-assisted project creation/modification, the complete installed plugin universe, governed remote wp-admin operations, themes/blocks/layout, third-party integrations and production-safe changes. Use when the user requests wordpress expert or invokes wordpress-expert. Do NOT use for unrelated languages, domains or production deployment.
whatsapp-simulator
Generate synthetic WhatsApp-style conversation fixtures for testing conversational systems, support flows and automation. Use when the user requests whatsapp simulator or invokes whatsapp-simulator. Do NOT use for unrelated tools or unapproved remote writes.
web-quality-audit
Assimilated TLC/web-quality multi-area audit. Review performance, accessibility, SEO and best practices in one pass. Triggers: audit my site, review web quality, check page quality, optimize my website across areas. Do not use for a single-area job (core-web-vitals, ai-seo, best-practices, perf-lighthouse).
voiceai
Provide an evidence-backed voice interface for AEOS. Activating this skill is treated as explicit current-session consent for `voiceai` to record user audio or transcribe audio from user-authorized audio/video media within the declared scope. The skill can: Use when the user requests voiceai or invokes voiceai. Do NOT use for unrelated tools or unapproved remote writes.