overthinkos/unsloth-studio
ov-images/skills/unsloth-studio/SKILL.md
Unsloth Studio fine-tuning web UI with CUDA GPU support, vLLM inference, and llama.cpp. Runs as a supervisord service on ports 8888 (Studio) and 8000 (vLLM API). MUST be invoked before building, deploying, configuring, or troubleshooting the unsloth-studio image.
development
Updated Apr 24, 2026
$ install --global
skillsauthnpx skillsauth add overthinkos/overthink-plugins unsloth-studioInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
Security Scan Results
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
3
Scanners Passed
9
Scanners in report
Clean
TrivyContainer and dependency vulnerability scanner
95%
Clean
SemgrepStatic code analysis for vulnerabilities
95%
Clean
mcp-scan (Snyk)Model Context Protocol security validation
95%
Skipped
Snyk (dep)Open source security scanning
50%
Skipped
Socket.devSupply chain security analysis
50%
Skipped
VirusTotalMulti-engine malware detection
50%
Skipped
CrowdStrikeAdvanced threat intelligence
50%
Skipped
OSV-ScannerOpen Source Vulnerability database check
50%
Skipped
OWASP Dep-Check
50%
Last scanned: Apr 24, 2026, 2:31 PM140.4s1 file scanned
SKILL.md
- name:
- unsloth-studio
- description:
- |
unsloth-studio
Unsloth Studio web UI for LLM fine-tuning with GPU acceleration.
Image Properties
| Property | Value | |----------|-------| | Base | nvidia | | Layers | agent-forwarding, unsloth-studio, notebook-finetuning, dbus, ov | | Platforms | linux/amd64 | | Ports | 8888, 8000 | | Registry | ghcr.io/overthinkos |
Layer Composition
The unsloth-studio layer is a Tier 2 environment-owner meta-layer that:
- Owns the pixi.toml (fine-tuning Python environment)
- Composes two Tier 1 sub-layers via
layers: [llama-cpp, unsloth] - Defines the supervisord service for the Studio web UI
Build order: pixi environment → llama-cpp (binaries) → unsloth (vLLM 0.19 wheel + unsloth pip + torch.compile patch) → supervisord config
Full Layer Stack
fedora→nvidia(CUDA base)pixi→python→supervisord(transitive)unsloth-studio— Tier 2 meta-layer (owns pixi.toml, service config)llama-cpp— llama.cpp binaries (Tier 1, vialayers:)unsloth— vLLM 0.19 + unsloth pip install + torch.compile patch (Tier 1, vialayers:)
Ports
| Port | Service | Protocol | |------|---------|----------| | 8888 | Unsloth Studio UI | HTTP | | 8000 | vLLM API server | HTTP |
Volumes
| Name | Path | Purpose | |------|------|---------| | models | ~/.cache/huggingface | HuggingFace model cache | | workspace | ~/workspace | Training data and outputs |
Quick Start
ov image build unsloth-studio
ov config unsloth-studio
ov start unsloth-studio
# Open http://localhost:8888
Key Layers
/ov-layers:unsloth-studio— Studio web UI service + pixi.toml (Tier 2)/ov-layers:llama-cpp— llama.cpp binaries (Tier 1 sub-layer)/ov-layers:unsloth— vLLM 0.19 + unsloth fine-tuning + torch.compile patch (Tier 1 sub-layer)/ov-layers:notebook-finetuning— 37 Unsloth fine-tuning notebooks provisioned into workspace volume/ov-layers:nvidia— GPU runtime and CDI device auto-detection (base)/ov-layers:cuda— CUDA toolkit and libraries (via nvidia base)/ov-layers:dbus— session bus for desktop notifications/ov-layers:ov— in-containerovbinary (enablesov test dbus notify)/ov-layers:agent-forwarding— SSH/GPG/direnv agent forwarding
Related Images
/ov-images:nvidia— parent (GPU without Studio)/ov-images:jupyter-ml— alternative ML UI with JupyterLab + CRDT MCP (same Tier 1 sub-layers)/ov-images:python-ml— ML libraries without any UI/ov-images:jupyter— legacy Jupyter with ML (shares port 8888)
Verification
After ov start:
ov status unsloth-studio— container runningov service status unsloth-studio— all services RUNNINGcurl -s -o /dev/null -w '%{http_code}' http://localhost:8888— Studio HTTP returns 200
When to Use This Skill
MUST be invoked when the task involves the unsloth-studio image, LLM fine-tuning via web UI, or Unsloth Studio deployment. Invoke this skill BEFORE reading source code or launching Explore agents.
Related
/ov:image— image family umbrella (image:entries inoverthink.yml, build/validate/inspect/list)/ov:build—build.ymlvocabulary (distros, builders, init-systems)
Related Skills
overthinkos/agents
development
VerifiedTrustedCommunity
Claude Code multi-agent support in Overthink — sub-agents, dynamic workflows, and agent teams, and how each drives the existing `ov eval` disposable beds to test and verify. MUST be invoked before authoring or invoking an ov sub-agent / dynamic workflow / agent team, wiring agent-lifecycle hooks, or asking "which primitive should drive the R10 beds?".
SKILL.mdUpdated May 31, 2026
overthinkos/workspace-mount
tools
VerifiedTrustedCommunity
Mounts a virtiofs share tagged `workspace` at /workspace inside a VM guest via a systemd .mount unit. Use when a kind:vm entity shares a host directory into the guest and you need it auto-mounted (and re-mounted at every boot).
SKILL.mdUpdated May 30, 2026
overthinkos/android
development
VerifiedTrustedCommunity
MUST be invoked before any work involving: the `kind: android` schema kind, a `target: android` deploy, the `apk:` layer package format (installing Android apps declaratively), AndroidDeployTarget, an in-pod emulator OR a remote/physical adb-endpoint device, or nested `pod → android` deployment. The first-class Android device + app surface that sits above `ov eval adb`/`appium`.
SKILL.mdUpdated May 26, 2026
overthinkos/git-workflow
tools
VerifiedTrustedCommunity
Use when committing, branching, pushing, merging, tagging, creating PRs, or approving/merging PRs with gh — the feat/-branch, R10-gated, never-force-push landing workflow across the main repo + the plugins submodule + image/<distro> submodules. Covers sync-to-upstream, branch/worktree pruning, the fork+PR path for contributors without write access, and cross-repo @github landing order.
SKILL.mdUpdated May 25, 2026