unsloth-studio-training

name	unsloth-studio-training
description	Manages LLM fine-tuning through Unsloth Studio API. Use when user wants to train models, check training status, manage datasets, export trained models, or work with Docker setup for Unsloth Studio. Triggers: fine-tune, training, dataset, export, GGUF, LoRA, QLoRA.

Unsloth Studio Training

Manage LLM fine-tuning jobs through Unsloth Studio's local web UI and API.

Quick Start

# Check if studio is running
curl -s http://localhost:8888/api/health | jq

# Login and get token
RESPONSE=$(curl -s -X POST http://localhost:8888/api/auth/login \
  -H "Content-Type: application/json" \
  -d '{"username": "admin", "password": "your-password"}')
export UNSLOTH_STUDIO_TOKEN=$(echo $RESPONSE | jq -r '.access_token')

# Start training
curl -X POST http://localhost:8888/api/train/start \
  -H "Authorization: Bearer $UNSLOTH_STUDIO_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model_name": "unsloth/llama-3.2-3b-instruct-bnb-4bit", ...}'

Workflow

1. Docker Setup     → [docker.md](references/docker.md)
2. Authenticate     → [authentication.md](references/authentication.md)
3. Prepare Dataset  → [datasets.md](references/datasets.md)
4. Train Model      → [training.md](references/training.md)
5. Export Model     → [export.md](references/export.md)
6. Run Inference    → [inference.md](references/inference.md)

Reference Files

File	Purpose
docker.md	Docker discovery, setup, container management
authentication.md	JWT auth, login/signup, token refresh
training.md	Start/stop jobs, monitor progress, hyperparams
datasets.md	Upload, list, format datasets
export.md	GGUF, merged, LoRA export, HF Hub
inference.md	Chat completion, streaming, tools
data-recipes.md	Dataset creation workflow
cli.md	Command-line reference

API Endpoints

Method	Endpoint	Description
GET	/api/health	Health check
POST	/api/auth/login	Login
POST	/api/train/start	Start training
GET	/api/train/stream	SSE progress stream
POST	/api/train/stop	Stop training
GET	/api/models/	List models
GET	/api/datasets/	List datasets
POST	/api/inference/chat	Chat inference

Training Methods

Method	VRAM (7B)	Description
qlora	~4GB	4-bit quantized + LoRA
lora	~16GB	Full precision + LoRA
full	~28GB	All weights trained

Environment Variables

UNSLOTH_STUDIO_URL=http://localhost:8888
UNSLOTH_STUDIO_TOKEN=<jwt_token>

Related Docs

• Unsloth Studio Docs
• Getting Started
• Data Recipes