feat(unsloth-training): add plugin manifest fields and marketplace entry

Add license, claude.minVersion, capabilities, triggers, and skills
array to plugin.json following established plugin conventions.
Register plugin in marketplace.json.

Co-Authored-By: duyetbot <duyetbot@users.noreply.github.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: claude

marketplace.json

@@ -114,11 +114,19 @@
       "version": "1.0.0",
       "type": "skill",
       "category": "monitoring"
+    },
+    {
+      "name": "unsloth-training",
+      "id": "unsloth-training@duyet-claude-plugins",
+      "description": "Fine-tune LLMs with Unsloth using GRPO or SFT. Dataset prep, synthetic data, environment flags, MLX (Apple Silicon), FP8, vision, mobile export, GGUF.",
+      "version": "1.0.0",
+      "type": "skill",
+      "category": "ml"
     }
   ],
   "metadata": {
     "lastUpdated": "2026-03-05",
-    "totalPlugins": 13,
-    "categories": ["agents", "git", "design", "productivity", "ui", "docs", "database", "monitoring"]
+    "totalPlugins": 14,
+    "categories": ["agents", "git", "design", "productivity", "ui", "docs", "database", "monitoring", "ml"]
   }
 }

unsloth-training/.claude-plugin/plugin.json

@@ -0,0 +1,36 @@
+{
+  "name": "unsloth-training",
+  "description": "Fine-tune LLMs with Unsloth using GRPO or SFT. Includes dataset preparation, synthetic data generation, environment flags, MLX (Apple Silicon), FP8, vision models, mobile export, GGUF.",
+  "version": "1.0.0",
+  "author": {
+    "name": "duyet"
+  },
+  "license": "MIT",
+  "claude": {
+    "minVersion": "1.0.0"
+  },
+  "capabilities": [
+    "skills"
+  ],
+  "triggers": [
+    "unsloth",
+    "fine.?tune",
+    "grpo",
+    "sft.*train",
+    "reward.*function",
+    "gguf",
+    "lora.*train",
+    "fp8",
+    "vision.*train",
+    "mlx.*train",
+    "qlora",
+    "mobile.*deploy"
+  ],
+  "skills": [
+    {
+      "name": "unsloth-training",
+      "path": "skills/unsloth-training/SKILL.md",
+      "description": "Fine-tune LLMs with Unsloth using GRPO or SFT. Supports FP8, vision models, mobile deployment, Docker, packing, and synthetic data generation."
+    }
+  ]
+}

unsloth-training/README.md

@@ -0,0 +1,114 @@
+# Unsloth Training Plugin
+
+Fine-tune LLMs with [Unsloth](https://unsloth.ai/) using GRPO (reinforcement learning) or SFT (supervised fine-tuning).
+
+## Install
+
+```bash
+/plugin install unsloth-training@duyet-claude-plugins
+```
+
+Or via [skills.sh](https://skills.sh):
+```bash
+npx skills add duyet/claude-plugins/unsloth-training
+```
+
+## Features
+
+- **GRPO** — RL with reward functions (no labeled outputs needed)
+- **SFT** — Supervised fine-tuning with input/output pairs
+- **Vision** — VLM fine-tuning (Qwen3-VL, Gemma3, Llama 3.2 Vision)
+- **FP8 Training** — 60% less VRAM, 1.4x faster (RTX 40+, H100)
+- **Packing** — 2-5x speedup for mixed-length data
+- **MLX** — Apple Silicon training via mlx-tune
+- **Export** — GGUF, Ollama, vLLM, LM Studio, SGLang
+- **Dataset Prep** — ChatML/ShareGPT/Alpaca formats, synthetic data generation
+- **Mobile** — QAT + ExecuTorch for iOS/Android deployment
+
+## Skills
+
+| Skill | Description |
+|-------|-------------|
+| `unsloth-training` | Core training guide with GRPO, SFT, reward design, model selection, export |
+
+## Reference Docs
+
+| Reference | Description |
+|-----------|-------------|
+| `installation.md` | pip/uv install, CUDA versions, venv, Colab |
+| `datasets-guide.md` | Dataset formats, chat templates, synthetic data |
+| `environment-flags.md` | Unsloth env flags (RETURN_LOGITS, COMPILE_DISABLE, etc.) |
+| `mlx-training.md` | Apple Silicon training with mlx-tune |
+| `fp8-training.md` | FP8 setup, VRAM savings |
+| `reward-design.md` | Reward function patterns for GRPO |
+| `domain-examples.md` | Voice AI, Sales Agent, Support examples |
+| `hyperparameters.md` | GRPOConfig/SFTConfig reference |
+| `export-formats.md` | GGUF, Ollama, vLLM, Dynamic 2.0 |
+| `advanced-training.md` | 500K context, packing, checkpoints |
+| `vision-training.md` | VLM fine-tuning |
+| `mobile-deployment.md` | QAT, ExecuTorch, iOS/Android |
+| `deployment.md` | Docker, vLLM, LoRA hot-swap, SGLang |
+| `troubleshooting.md` | Common fixes |
+
+## Usage Examples
+
+```
+# Natural language triggers
+"fine-tune Qwen3.5 for classification"
+"train with GRPO and reward functions"
+"prepare dataset in ChatML format"
+"install unsloth on Colab"
+"train on Apple Silicon with MLX"
+"export model to GGUF for Ollama"
+```
+
+## Quick Start
+
+### SFT
+```python
+from unsloth import FastLanguageModel
+from trl import SFTTrainer, SFTConfig
+
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name="unsloth/Qwen3.5-4B",
+    max_seq_length=512,
+    load_in_4bit=False,
+    load_in_16bit=True,
+)
+model = FastLanguageModel.get_peft_model(model, r=16)
+
+trainer = SFTTrainer(
+    model=model,
+    train_dataset=dataset,
+    processing_class=tokenizer,
+    args=SFTConfig(
+        per_device_train_batch_size=2,
+        num_train_epochs=3,
+        learning_rate=2e-4,
+        packing=True,
+    ),
+)
+trainer.train()
+model.save_pretrained_gguf("model", tokenizer, quantization_method="q4_k_m")
+```
+
+### GRPO
+```python
+from trl import GRPOConfig, GRPOTrainer
+
+def correctness_reward(completions, answer, **kwargs):
+    return [2.0 if extract_answer(c) == a else 0.0
+            for c, a in zip(completions, answer)]
+
+trainer = GRPOTrainer(
+    model=model,
+    args=GRPOConfig(num_generations=4, beta=0.04, learning_rate=5e-6),
+    train_dataset=dataset,
+    reward_funcs=[correctness_reward],
+)
+trainer.train()
+```
+
+## License
+
+Apache-2.0