overthinkos/notebook-finetuning
jupyter/skills/notebook-finetuning/SKILL.md
Unsloth fine-tuning notebook collection provisioned into the workspace volume at deploy time. Data-only candy — no packages, no services, no dependencies. Use when working with notebook-finetuning, Unsloth training notebooks, or unsloth-studio data provisioning.
testing
Updated Jun 14, 2026
$ install --global
skillsauthnpx skillsauth add overthinkos/overthink-plugins notebook-finetuningInstall 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: Jun 14, 2026, 5:58 AM127.0s1 file scanned
SKILL.md
- name:
- notebook-finetuning
- description:
- |
notebook-finetuning -- Unsloth fine-tuning notebook data candy
Candy Properties
| Property | Value |
|----------|-------|
| Dependencies | (none) |
| Packages | (none) |
| Services | (none) |
| Volumes | workspace -> /workspace (from unsloth-studio) |
| Data | data/finetuning -> workspace volume, dest: finetuning |
| Install files | (none) |
How It Works
This is a data candy — it uses the data: field in charly.yml to map a directory of notebooks to a named volume with a subdirectory destination:
info: "Unsloth fine-tuning notebook collection for unsloth-studio"
data:
- src: data/finetuning
volume: workspace
dest: finetuning
At build time, the contents of data/finetuning/ are staged into /data/workspace/finetuning/ inside the box.
At deploy time, when the workspace volume is configured as a bind mount (charly config --bind workspace), charly config copies the staged data into the host-backed volume directory at <workspace>/finetuning/. This seeds the volume with ready-to-use training notebooks.
The dest: finetuning field places the notebooks in a subdirectory rather than the volume root, keeping the workspace organized alongside other content.
Included Data
37 Jupyter notebooks + 1 manifest, organized by training category:
| Category | Notebooks | Models |
|----------|-----------|--------|
| 00-Setup | 00_Unsloth_Setup.ipynb | (general) |
| 01-FastInference | 3 notebooks | Llama, Qwen, Qwen_Think |
| 02-Vision Training | 2 notebooks | Ministral, Pixtral |
| 03-SFT Training | 5 notebooks | Ministral (text + vision), Pixtral, Qwen, Qwen_Think |
| 04-GRPO Training | 5 notebooks | Ministral (text + vision), Pixtral, Qwen, Qwen_Think |
| 05-DPO Training | 2 notebooks | Qwen, Qwen_Think |
| 06-Reward Training | 2 notebooks | Qwen, Qwen_Think |
| 07-RLOO Training | 5 notebooks | Ministral (text + vision), Pixtral, Qwen, Qwen_Think |
| 08-QLoRA | 12 notebooks | Ministral, Qwen_Think (alpha scaling, continual learning, rank comparison, multi-adapter, quantization comparison, target modules) |
Manifest: notebooks.yaml — structured catalog of all notebooks with metadata.
Usage
# charly.yml
unsloth-studio:
candy:
- unsloth-studio
- notebook-finetuning
# ... other candies
# Deploy with bind-backed workspace volume
charly config unsloth-studio --bind workspace
# Notebooks are seeded at <workspace>/finetuning/ on first config
charly start unsloth-studio
# Open http://localhost:8888 → navigate to finetuning/
Notebook Compatibility Fixes
The notebooks include several workarounds for upstream library changes:
packing=Truein SFTConfig (19 notebooks) -- Required for TRL 1.0 compatibility. Without it, SFT training fails with the updated TRL APIos.environ["UNSLOTH_ENABLE_FLEX_ATTENTION"] = "0"(16 Ministral/Pixtral notebooks) -- Disables flex_attention to work around a transformers 5.5 bug that crashes these model architecturesmax_memory={0: "14GB"}in model loading (3 Pixtral-12B notebooks) -- Fixes accelerate device_map estimation for Pixtral-12B models that would otherwise OOMmax_prompt_lengthremoved from DPOConfig (2 DPO notebooks) -- Parameter deprecated and removed in TRL 1.0
Used In Boxes
/charly-jupyter:unsloth-studio/charly-jupyter:jupyter-ml-notebook
Related Skills
/charly-image:layer-- data field documentation and candy authoring rules/charly-core:charly-config-- data provisioning duringcharly configsetup/charly-core:deploy-- volume backing configuration (bind, named, encrypted)/charly-jupyter:unsloth-studio-- the Tier 2 meta-layer that owns the workspace volume/charly-jupyter:notebook-templates-- sibling data candy pattern (starter notebooks for jupyter)/charly-jupyter:unsloth-studio-- the box that includes this candy
When to Use This Skill
Use when the user asks about:
- The notebook-finetuning candy or its contents
- Unsloth fine-tuning notebook templates or training workflows
- How training notebooks get provisioned into the workspace volume
- The
data:field withdest:subdirectory mapping - Which models and training methods are covered by the notebook collection
Related
/charly-check:check— declarative testing (check:block,charly check box,charly check live)
Related Skills
overthinkos/plugin
tools
VerifiedTrustedCommunity
Use when authoring or modifying a charly PLUGIN — a candy with a `plugin:` block that contributes Providers (verbs/kinds/deploy-targets/steps/builders/commands), its own CUE schema, builtin (compiled-in) or external (out-of-tree git repo). Covers the unified Provider model, the per-plugin CUE-schema contract (single source → Go params for dev + schema-over-Describe RPC for runtime), the SDK, and the loader.
SKILL.mdUpdated Jun 22, 2026
overthinkos/cue
tools
VerifiedTrustedCommunity
The CUE data-validation / configuration CLI (cue), pinned to v0.16.1. Use when working with the cue candy, installing the cue binary into a box or onto a target:local dev host, or running the offline schema-vendoring pipeline that feeds charly's egress validation.
SKILL.mdUpdated Jun 16, 2026
overthinkos/egress
tools
VerifiedTrustedCommunity
CUE EGRESS validation — validating (and, where it adds value, generating) the config files charly WRITES to a system BEFORE the bytes hit disk. MUST be invoked before working on charly/egress.go, the vendored schemas under candy/plugin-egress/egress-schemas/vendor/, the ValidateEgress / registerVendoredEgressKind path, the offline `task cue:vendor` pipeline, or adding an egress schema for any written artifact (cloud-init, k8s manifests, traefik routes, runtime config, install ledger, systemd/quadlet units, ssh_config, libvirt XML).
SKILL.mdUpdated Jun 16, 2026
overthinkos/check-k8s
tools
VerifiedTrustedCommunity
Kubernetes cluster-probe declarative check verb — the `kube:` check verb (nodes, pods, ingress, storage class, addon health, apply/delete, and arbitrary resource GETs) served out-of-process by the candy/plugin-kube plugin (vendored client-go; no external kubectl required).
SKILL.mdUpdated Jun 14, 2026