Fine-tune LLMs up to 2x faster with 70% less VRAM. Web UI for training (LoRA, full fine-tune), inference, data recipes, and model export. OpenAI-compatible API included.
| Port | Service | URL Pattern |
|---|---|---|
| 8000 | Unsloth Studio | https://<POD_ID>-8000.proxy.runpod.net |
| 8888 | JupyterLab | https://<POD_ID>-8888.proxy.runpod.net |
| 22 | SSH | Via Runpod SSH connection |
| Variable | Required | Description |
|---|---|---|
UNSLOTH_ADMIN_PASSWORD |
Yes | Admin password for Unsloth Studio |
JUPYTER_PASSWORD |
No | Token for JupyterLab access (empty = no auth) |
EXIT_ON_CRASH |
No | Set to true to exit container on Studio crash (default: false, keeps container alive for debugging) |
All persistent data is stored under /workspace/unsloth/:
cache/huggingface/— Downloaded model weightsoutputs/— Training checkpointsexports/— Exported models (GGUF, safetensors)assets/datasets/— Uploaded and generated datasetsauth/— User authentication databaselogs/— Service logs (Jupyter)
With a network volume attached, all data persists across pod restarts.
| Field | Value |
|---|---|
| Exposed HTTP Ports | 8000,8888 |
| Exposed TCP Ports | 22 |