2.3.51 Satellite Unsloth

Unsloth

Handle: unsloth
URL: http://localhost:34511

Unsloth is a Jupyter Lab environment pre-configured with the Unsloth library for fast LLM fine-tuning. It provides 2x faster training with 70% less memory usage compared to standard fine-tuning methods, along with pre-loaded example notebooks for various fine-tuning tasks.

The service includes:

• Jupyter Lab with all dependencies pre-installed
• Unsloth library for efficient LLM fine-tuning
• Pre-configured notebooks for supervised fine-tuning, reinforcement learning (GRPO, DPO, ORPO, KTO), and vision fine-tuning
• SSH access for remote development
• GPU support with NVIDIA Container Toolkit

Starting

# Pull the pre-built image
# Image is ~7GB which is typical for CUDA-enabled environments
harbor pull unsloth

# Start the service (opens Jupyter Lab in browser)
harbor up unsloth --open

Access Jupyter Lab at http://localhost:34511 with the password (default: unsloth, configurable).

Configuration

Environment Variables

Following options can be set via harbor config:

# Jupyter Lab Configuration
HARBOR_UNSLOTH_HOST_PORT=34511              # Jupyter Lab web interface port
HARBOR_UNSLOTH_SSH_PORT=34512               # SSH access port (optional)
HARBOR_UNSLOTH_JUPYTER_PASSWORD="unsloth"   # Jupyter Lab password
HARBOR_UNSLOTH_WORKSPACE="./unsloth/workspace"  # Local workspace directory

# Docker Image
HARBOR_UNSLOTH_IMAGE="unsloth/unsloth"      # Docker image to use
HARBOR_UNSLOTH_VERSION="latest"             # Image version/tag

Volumes

The service mounts the following directories:

• HARBOR_UNSLOTH_WORKSPACE → /workspace/work - Your working directory for notebooks and data
• HARBOR_HF_CACHE → /root/.cache/huggingface - HuggingFace model cache

Usage

Accessing Pre-loaded Notebooks

After starting the service, navigate to the unsloth-notebooks folder in Jupyter Lab to access example notebooks for:

• Text fine-tuning (Llama, Qwen, Gemma, Mistral, etc.)
• Reinforcement learning (GRPO, DPO, ORPO, KTO, SimPO)
• Vision fine-tuning (Qwen-VL, LLaVA)
• Text-to-speech fine-tuning
• Quantization-aware training

HuggingFace Token

To download models from HuggingFace Hub, set your token:

harbor config set hf.token "hf_your_token_here"

Examples

Basic Fine-tuning Workflow

1. Start Unsloth and open Jupyter Lab
2. Navigate to unsloth-notebooks/
3. Choose a notebook (e.g., Llama3.1_(8B)-Alpaca.ipynb)
4. Run all cells or customize for your dataset
5. Export your fine-tuned model to GGUF, Ollama, vLLM, or HuggingFace

Using Your Own Data

Place your datasets in the workspace directory (mapped to /workspace/work):

# From your local machine
cp my_dataset.json $(harbor home)/services/unsloth/workspace/

# In Jupyter Lab, access it at:
# /workspace/work/my_dataset.json

Related Services

• Jupyter - Standard Jupyter Lab environment
• vLLM - For serving fine-tuned LLMs