FD unsloth-fine-tuning
Fast LLM fine-tuning with Unsloth - 2-5x faster training, 50-80% less VRAM. Use for single-GPU LoRA/QLoRA SFT, GRPO/RL reasoning training, vision/TTS fine-tuning, and GGUF export to Ollama/vLLM/llama.cpp. Supports 300+ models including Llama, Qwen, Gemma, DeepSeek, Mistral, Phi, and gpt-oss.
Fast LLM fine-tuning with Unsloth - 2-5x faster training, 50-80% less VRAM.
As a process D 43/100 · Unfinished process — weak spots: result and completion, inputs and preconditions, consistency
What is at stake
The skill contains fragments that, in the wrong hands, cost money or data. Below: what the installer risks and what the author should do.
The instructions or scripts send data out: environment variables, keys, file contents, chat history. You may never notice, because the agent performs the upload as "part of the task".
If the upload is not needed for the task, remove it: catalogs flag such skills and delist them. If it is needed, name the destination explicitly, say what leaves the machine, and give the user a switch.
The skill contains commands that delete files, rewrite disks or execute code fetched from the network. The agent may run them without asking if it believes the instructions require it.
Replace destructive commands with safe equivalents that ask for confirmation, scope them to one folder, and stop piping curl into a shell: pin a version and a checksum.
How to improve
- Address the high-severity findings: each costs 18 safety points. If one is a false positive, add the rule id to guard.allow in spec.yaml.
- For Hermes the description must be one sentence under 60 characters; move the conditions to a "When to Use" section.
- The SKILL.md body is over 5,000 tokens: move reference detail into references/ and load it when needed.
- Your own cases (evals/evals.json, 4–6 real requests with expected answers): the full check would then run those instead of a model-drafted suite.
- A spec.yaml with trigger phrases and assertions — a behaviour contract for CI; `skilltest init` writes a template.
Guard findings · 5
-
high Exfiltration
exfil-send-secrets-to-urldocs/saving-to-ollama.md:7Instruction to send secrets/history to an external endpoint (destination is a well-known publishing service)You can save the finetuned model as a small 100MB file called a LoRA adapter. You can instead push to the Hugging Face hub as well if you want to upload your model! Remember to get a Hugging Face toke
known service -
high Dangerous commands
cmd-pipe-to-shelldocs/sglang-guide.md:15Downloads and executes remote code from an unrecognised host (pipe to shell)curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
-
high Exfiltration
exfil-send-secrets-to-urldocs/tutorial-llama3-ollama.md:284Instruction to send secrets/history to an external endpoint (destination is a well-known publishing service)We can now save the finetuned model as a small 100MB file called a LoRA adapter like below. You can instead push to the Hugging Face hub as well if you want to upload your model! Remember to get a Hug
known service
Medium and low: 2
-
medium Dangerous commands
cmd-pipe-to-shell-known-hostdocs/qwen3-coder-next.md:583Pipe-to-shell installer from a well-known host (still executes remote code)curl -LsSf https://astral.sh/uv/install.sh | sh
-
low Secrets in code
secret-high-entropy-tokendocs/datasets.md:268High-entropy token-like string (may be an id, hash or a credential)If you want to fine-tune a model that already has reasoning capabilities like the distilled versions of DeepSeek-R1 (e.g. Deep…-8B), you will need to still follow question/task and
Files scanned: 41. Evidence is masked. Grey chips explain why severity was lowered.
Against the Agent Skills spec
- warning
description-long-hermesdescription is 292 chars; the Hermes authoring standard requires ≤ 60 (one sentence, ending with a period) - warning
body-longSKILL.md body ≈ 6446 tokens (recommended < 5000); move details to references/ - note
frontmatter-keyunknown frontmatter key "dependencies"
Process rating: all ten parameters 43/100
- 0Result and completion. Does not say what the result is
- 0Inputs and preconditions. Does not say what the process needs to start
- 0Progress reporting. Says nothing while it works
- 30Running it twice. 12 mutating operations with no state check
- 40Consistency. Frontmatter name (unsloth-fine-tuning) differs from the folder (unsloth)
- 50When it triggers. No condition that starts the skill
- 50Failures and branches. 0 branches, has a failure section
- 60Tools and files. Uses tools (python, node) that frontmatter does not declare
- 70Execution cost. Instruction body is 6446 tokens
- 85Steps. 81 steps, 2 vague phrases
- low 16 top-level sections: this looks like several domains in one skill
- low The response is described with custom markup (5 tags): a typed call is more reliable
Everything here is measured from the skill text rather than judged by a model, so the numbers are checkable. A parameter weighs more when it is a more common reason for the process to stall.
Quality signals
- +5Description has no quoted example phrases that should trigger the skill
- +3Output format is not stated: the model decides each time
- +2Single-language instructions
- +4Description says when NOT to use the skill
- +3Description length 292: enough signal without eating the budget
- +4Structure: 46 headings
- +3Step-by-step instructions: 81 items
- +4Has examples (10 code blocks)
- +4Reference files are cited in the instructions (29 of 40)
- +1License stated
Quality base 70; lint remarks subtract, signals add up to 100. Result: 76.