mint-lora-training-skills

AI Workflow Skills

Reusable agent skills for AI training and fine-tuning workflows.

English · 中文文档 · 日本語

Install

Download a released skill zip from the links below, then extract it into your agent's skill directory:

unzip grpo-finetune-2.0.0.zip -d ~/.codex/skills/
unzip mint-lora-training-1.0.0.zip -d ~/.codex/skills/

Claude Code users can also install this repository as a plugin marketplace source when plugin marketplace support is available:

/plugin marketplace add Taiyi-AI-Lab/AI-Workflow-Skills

Skills

grpo-finetune

Category: AI Training / Fine-Tuning

Best for: running an end-to-end local GRPO fine-tuning workflow for LLMs, including environment checks, reward functions, training script templates, dry runs, full training, evaluation, and troubleshooting.

Highlights:

• 7-step GRPO workflow from requirements gathering to final evaluation
• Ready-to-run templates for rewards, training, and evaluation
• Version-sensitive notes for TRL, vLLM, CUDA, OOM handling, and batching
• References for reward design, training-curve interpretation, and debugging

Links: README · SKILL.md · Download v2.0.0 .zip

mint-lora-training

Category: AI Training / Fine-Tuning

Best for: running cloud-hosted MinT LoRA fine-tuning on macaron.im without local GPU training, including SFT, DPO, GRPO, evaluation, adapter download, and quota checks.

Highlights:

• Clear cloud-vs-local boundaries for MinT training
• Ready-to-run scripts for setup, SFT, DPO, GRPO, evaluation, download, and quota
• Workflow guides for first run, supervised fine-tuning, preference tuning, and RL
• Reference notes for supported models, endpoints, data formats, and known pitfalls

Links: README · SKILL.md · Download v1.0.0 .zip

Maintainer Commands

npm run list
npm run pack -- --skill grpo-finetune
npm run pack -- --skill mint-lora-training
npm run validate

npm run validate checks manifests, builds release zips, verifies README download markers, compiles bundled Python scripts, and scans for secrets or private local paths.

Release

Per-skill releases use this tag format:

grpo-finetune-v2.0.0
mint-lora-training-v1.0.0

Use npm run release:dry before publishing. The tag-triggered GitHub Action validates the manifest version, builds a zip and sha256 file, creates a GitHub Release, and syncs the README download links.

Layout

skills/
├── grpo-finetune/
│   ├── README.md
│   ├── SKILL.md
│   ├── manifest.json
│   ├── references/
│   └── templates/
└── mint-lora-training/
    ├── README.md
    ├── SKILL.md
    ├── manifest.json
    ├── scripts/
    └── workflows/