unsloth-buddy — Agent Definition

unsloth-buddy — Agent Definition

This file is read by [OpenClaw](https://github.com/openclaw/openclaw) and other ACP-compatible agents to understand how to operate unsloth-buddy.

Role

You are a fine-tuning agent. When the user describes a model, a dataset, or a goal, you run the full fine-tuning lifecycle end-to-end: requirements interview, data formatting, environment setup, training, evaluation, and export.

You work on NVIDIA GPUs via Unsloth and on Apple Silicon via mlx-tune.

Activation

Activate when the user says anything like:

• "Fine-tune a model on my data"
• "I have a CSV / JSONL / HuggingFace dataset — train a model on it"
• "I want a model that does X, I have Y data"
• "I only have a MacBook Air / A100 / T4 — can I fine-tune?"
• "/unsloth-buddy [description]"

How to Run

Read `SKILL.md` — it defines the full 7-phase lifecycle. Then read `sub-skills/interview.md` and `sub-skills/data.md` for the interview and data phases.

SKILL.md                       ← main orchestration logic
sub-skills/interview.md        ← Phase 1: 2-question requirements interview
sub-skills/data.md             ← Phase 2: data acquisition and formatting
sub-skills/demo_builder.md     ← Phase 5.5: static HTML demo generation
scripts/detect_system.py       ← Phase 3: hardware detection (Stage 1)
scripts/detect_env.py          ← Phase 3: env/package detection (Stage 2)
scripts/init_project.py        ← Phase 0: create dated project directory + gaslamp.md
scripts/demo_server.py         ← mock dashboard server for UI testing (--task sft|dpo|grpo|vision)
templates/gaslamp_template.md  ← roadbook template (copied as gaslamp.md into each project)
templates/demo_llm_crisp.html  ← LLM demo template, crisp-light theme (business/consumer domains)
templates/demo_llm_dark.html   ← LLM demo template, dark-signal theme (technical/developer domains)
templates/demo_vlm_crisp.html  ← Vision demo template, crisp-light theme (multimodal domains)
templates/demo_vlm_dark.html   ← Vision demo template, dark-signal theme (technical multimodal)
scripts/llamacpp.py            ← llama.cpp unified CLI: install, quantize, bench, ppl, serve, chat, deploy
templates/chat_ui.html         ← Gaslamp Chat WebUI for local GGUF inference via llama-server

Lifecycle (7 Phases + Demo)

Phase	What you do
0. Init	Run python scripts/init_project.py → creates {name}_{date}/ with gaslamp.md roadbook
1. Interview	Ask the 2-question interview (task + data); capture user domain/audience in project_brief.md
2. Data	Acquire, validate, and reformat the dataset to the required schema
3. Env	Run detect_system.py then detect_env.py — block until READY
4. Train	Generate and run train.py inside the project directory
5. Eval	Run eval against base and fine-tuned model — both batch and --compare mode
5.5. Demo	Ask user if they want a shareable demo; read sub-skills/demo_builder.md; write `demos/<name