Get started with Unsloth Studio

Unsloth Studio is a local, browser-based GUI for fine-tuning LLMs without writing any code. It wraps the training pipeline in a clean interface that handles model loading, dataset formatting, hyperparameter configuration, and live training monitoring.

StudioData RecipeExportChatVideo

Setup Unsloth Studio

First, launch Unsloth Studio using either a local install or a cloud option. Follow the install instructions for your setup, or use our free Colab notebook. For a local setup, run:

uunsloth studio -H 0.0.0.0 -p 8888

Then open http://localhost:8888 in your browser.

On first launch you will need to create a password to secure your account and sign in again later.

You’ll then see a brief onboarding wizard to choose a model, dataset, and basic settings. You can skip it at any time and configure everything manually.

Studio - Quickstart

Unsloth Studio homepage has 4 main areas: Model, Dataset, Parameters, and Training/Config

• Easy setup for models and data from Hugging Face or local files

• Flexible training choices like QLoRA, LoRA, or full fine-tuning, with defaults filled in

• Helpful config tools for splits, column mapping, hyperparameters and YAML configs

• Great training visibility with live progress, GPU stats, charts, startup status

1. Select model and method

Model Type

Select the modality that matches your use-case:

Training Method

Three methods are available, toggled with a pill selector:

Type any Hugging Face model name or search the Hub directly from the combobox. Local models stored in ~/.unsloth/studio/models and your Hugging Face cache also appear in the list.

When you pick a model the Studio automatically fetches its configuration from the backend and pre-fills sensible defaults for all hyperparameters.

HuggingFace Token

Paste your Hugging Face access token here if the model is gated (e.g. Llama, Gemma). The token is validated in real-time and an error is shown inline if it is invalid.

2. Dataset

Source

Switch between two tabs to choose where your data comes from:

• HuggingFace Hub - live search against the Hub. The last-updated date is shown for each result.

• Local - drag-and-drop or click to upload a file unstructured or structured files like: PDF, DOCX, JSONL, JSON, CSV, or Parquet format. Previously uploaded datasets appear in a list that refreshes automatically.

Prompt Studio how to interpret and format your data:

Splits and Slicing

• Subset - automatically populated from the dataset card.

• Train split / Eval split - choose which splits to use. Setting an eval split enables the Eval Loss chart during training.

• Dataset slice - optionally restrict training to a row range (start index / end index) for quick experiments.

Column Mapping

If the Studio cannot automatically map your dataset columns to the correct roles a Dataset Preview dialog opens. It shows sample rows and lets you assign each column to instruction, input, output, image, etc. Suggested mappings are pre-filled where possible.

3. Hyperparameters

Parameters are grouped into collapsible sections. You can view our detailed LoRA hyperparameters guide here:

LoRA Settings

(Hidden when Full Fine-tuning is selected)

For Vision models with an image dataset, four additional checkboxes appear:

• Finetune Vision Layers

• Finetune Language Layers

• Finetune Attention Modules

• Finetune MLP Modules

Training Hyperparameters

Organized into three tabs:

4. Training and Config

The bottom-right card has three config management buttons and the Start Training button.

The Start Training button stays disabled until a model and dataset are both configured. Validation errors appear inline - for example, setting eval steps without choosing an eval split, or pairing a text-only model with a vision dataset.

Loading Screen

After you click Start Training, a full-page overlay appears while the backend prepares everything.

The overlay shows an animated terminal with live phase updates:

• Blue: Downloading model / dataset

• Amber: Loading model / dataset

• Blue: Configuring

• Green: Training

You can cancel at any time using the × button in the corner. A confirmation dialog will appear before anything is stopped.

Training Progress and Observability

Once the first training step arrives the overlay dismisses and the live training view is revealed. The fine-tuning process is complete when steps reach 100% on the progress bar. You can view the elapsed time and tokens.

Charts

Four live charts update as training progresses:

1. Training Loss - raw values plus an EMA-smoothed line and a running average reference line

2. Learning Rate - the LR schedule curve

3. Gradient Norm - gradient norm over steps

4. Eval Loss - only shown when you configured an eval split

Each chart has settings (gear icon) with:

Option

Default

Viewing window

Last N steps slider

EMA Smoothing

0.6

Show Raw

On

Show Smoothed

On

Show Average line

On

Scale (per series)

Linear / Log

Outlier clipping

No clip / p99 / p95

Config Files

All training configurations can be saved and reloaded as YAML files. Files are named automatically as:

The YAML is structured into three sections:

This makes it easy to reproduce runs, share configurations, or version-control your experiments.

Data Recipes - Quickstart

Unsloth Data Recipes lets you upload documents like PDFs or CSVs files and transforms them into useable datasets. Create and edit datasets visually via a graph-node workflow.

The recipes page is the main entry point. Recipes are stored locally in the browser, so you come back to saved work later. From here, you can create a blank recipe or open a guided learning recipe.

Data Recipes follows the same basic path. You open the recipes page, create or pick a recipe, build the workflow in the editor, validate it run a preview, then run the full dataset once the output looks right. Add seed data and generation blocks, validate the workflow, preview sample output, then run a full dataset build. Unsloth Data Recipes is powered by NVIDIA DataDesigner.

At a glance a usual workflow should look like this:

1. Open the recipes page.

2. Create a new recipe or open an existing one.

3. Add blocks to define your dataset workflow.

4. Click Validate to catch configuration issues early.

5. Run a preview to inspect sample rows quickly.

6. Run a full dataset build when the recipe is ready.

7. Review progress and output live in graph or in Executions view for mode details.

8. Select the resulting dataset in Studio and fine tune a model.

Export - Quickstart

Use Unsloth Studio 'Export' to export, save, or convert models to GGUF, Safetensors, or LoRA for deployment, sharing, or local inference in Unsloth, llama.cpp, Ollama, vLLM, and more. Export a trained checkpoint or convert any existing model.

You can read our detailed tutorial / guide about exporting models with Unsloth Studio here:

Chat - Quickstart

Unsloth Studio lets you run models 100% offline on your computer. Run model formats like GGUF and safetensors from Hugging Face or from your local files.

• Download + Run any model like GGUFs, fine-tuned adapters, safetensors etc.

• Compare different model outputs side-by-side

• Upload documents, images, and audio in your prompts

• Tune inference settings like: temperature, top-p, top-k and system prompt

You can read our detailed tutorial / guide about running models with Unsloth Studio here:

Video Tutorial

Here are 2 video tutorials to get you started with Unsloth Studio!

Advanced Settings

CLI Commands

The Unsloth CLI (cli.py) provides the following commands:

Project Structure

API Reference

All endpoints require a valid JWT Authorization: Bearer <token> header (except /api/auth/* and /api/health).