Unsloth Workflows

Use this skill as a framework guide and a shell workflow runner. Keep three modes distinct:

Discuss: explain concepts, explore models, or inspect a request. Covers Unsloth Studio/Desktop, cross-platform hardware, and agent runtimes. Do not install, write files, download models, or train.
Prepare: create reviewable configuration or task-specific scripts (e.g. custom RL reward functions, Data Recipes, multimodal pipelines). Do not install dependencies or start GPU work.
Execute: perform the requested operation after preflight. Installation and model downloads are side effects and must be limited to the selected project.

Route a request

Identify the objective: domain adaptation (CPT), supervised demonstrations (SFT), preferences (DPO/ORPO/KTO/SimPO), reinforcement learning (GRPO/GSPO/Agent RL), quantization-aware training (QAT), multi-token prediction (MTP), multimodal (vision/audio), embeddings, or serving/export.
Identify parameter method and precision: full fine-tuning, LoRA, QLoRA, FP8, or Dynamic NVFP4.
Inspect inputs and targets: model family/revision, chat/tool template, dataset schema, sequence length (up to 500K context), GPU memory, and target runtime (vLLM, GGUF, local API, Claude Code, Hermes Agent).
Select mode:
- For Discuss, use the references and live documentation queries without modifying disk state.
- For Prepare, author reviewable JSON configs and adaptation scripts.
- For Execute, run the runner commands below from this skill directory.