Is tune-fine-tuning-workflow safe to install?

SaferSkills independently audited tune-fine-tuning-workflow (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.

CategoryWeightCategory scoreContribution

Securityprompt, exec, net, exfil, eval

35%

100

35.0 pts

Supply chainhash, typosquat, maintainer, lockfile

20%

100

20.0 pts

Maintenancestaleness, pinning, CI

15%

100

15.0 pts

TransparencySKILL.md, perms, README

15%

100

15.0 pts

Communityinstalls, verify, response

15%

100

15.0 pts

Findings & checks · 0 flagged

Securityscore 100 · 0 findings

✓—All security checks passedNo findings in this category for the latest scan.pass

Supply chainscore 100 · 0 findings

✓—All supply chain checks passedNo findings in this category for the latest scan.pass

Maintenancescore 100 · 0 findings

✓—All maintenance checks passedNo findings in this category for the latest scan.pass

Transparencyscore 100 · 0 findings

✓—All transparency checks passedNo findings in this category for the latest scan.pass

Communityscore 100 · 0 findings

✓—All community checks passedNo findings in this category for the latest scan.pass

The primary manifest — the file an agent reads to learn what this artifact does.

solution-plays/13-fine-tuning-workflow/.github/skills/tune-fine-tuning-workflow/SKILL.md · 1 file

solution-plays/13-fine-tuning-workflow/.github/skills/tune-fine-tuning-workflow/SKILL.md4.9 KB · Markdown

name: tune-fine-tuning-workflow

description: "Tune Fine-Tuning — optimize LoRA rank, learning rate, epochs, batch size, data mix, compute selection, cost per training run. Use when: tune, optimize hyperparameters."

Tune Fine-Tuning Workflow

When to Use

Optimize LoRA/QLoRA hyperparameters for best quality
Balance training cost vs model quality
Select optimal compute (GPU type and count)
Configure data mixing strategies
Reduce cost per training run

Tuning Dimensions

Dimension 1: LoRA Hyperparameters

Parameter	Default	Range	Impact
`lora_rank`	16	4-128	Higher = more capacity, more VRAM, better quality
`lora_alpha`	32	2× rank	Controls effective learning rate scaling
`learning_rate`	1e-4	1e-5 to 5e-4	Too high = unstable, too low = slow convergence
`epochs`	3	1-10	More = better fit, risk of overfitting
`batch_size`	4	1-32	Larger = stable gradients, more VRAM
`warmup_steps`	100	0-500	Prevents early loss spikes
`weight_decay`	0.01	0-0.1	Regularization against overfitting
`target_modules`	q,v	q,k,v,o,gate,up,down	More modules = more capacity

Tuning strategy:

Start with defaults: rank=16, lr=1e-4, epochs=3
If underfitting: increase rank to 32-64, increase lr slightly
If overfitting: reduce epochs, add weight_decay, add more data
If training unstable: reduce lr, increase warmup_steps

Dimension 2: Data Quality Optimization

Factor	Current	Target	How to Improve
Sample count	10-50	200-500	Synthetic data augmentation
Diversity	Low	High	Cover all task categories
Quality	Mixed	Curated	Human review top 100 samples
Balance	Skewed	Balanced	Upsample minority categories
Length distribution	Fixed	Varied	Include short + long examples

Data augmentation techniques:

Paraphrase existing examples (GPT-4o with different wording)
Backtranslation (English → French → English)
Template variation (same content, different framing)
Never augment test/validation sets — only training data

Dimension 3: Compute Selection

Compute	GPU	VRAM	Best For	Cost/hr
Standard_NC6s_v3	1× V100	16GB	LoRA on 7B models	~$3.06
Standard_NC24ads_A100_v4	1× A100	80GB	LoRA on 7B-70B	~$3.67
Standard_ND96asr_v4	8× A100	640GB	Full fine-tune 70B+	~$27.20

Decision matrix:

LoRA on ≤7B model → NC6s_v3 (cheapest, sufficient VRAM)
LoRA on 13B-70B → NC24ads_A100_v4 (80GB single GPU)
QLoRA on 70B → NC24ads_A100_v4 (4-bit fits in 80GB)
Full fine-tune → ND96asr_v4 (8× A100 with DeepSpeed)

Dimension 4: Azure OpenAI Fine-Tuning Parameters

Parameter	Default	Range	Notes
`n_epochs`	auto	1-10	"auto" = Azure optimizes
`batch_size`	auto	1-256	"auto" = Azure optimizes
`learning_rate_multiplier`	1.0	0.1-10	Multiplied by base lr
`suffix`	—	string	Model name suffix

Cost: Azure OpenAI fine-tuning charges per 1K training tokens:

Model	Training Cost	Hosting Cost
gpt-4o-mini	$3.00/1M tokens	$0.30/1M input + $1.20/1M output
gpt-4o	$25.00/1M tokens	$3.75/1M input + $15.00/1M output

Dimension 5: Cost Per Training Run

Estimate for 500 samples, avg 1000 tokens, 3 epochs:

Method	Tokens	Cost
Azure OpenAI (gpt-4o-mini)	1.5M tokens	~$4.50
Azure OpenAI (gpt-4o)	1.5M tokens	~$37.50
Azure ML (LoRA on 7B, NC6s)	2 hrs compute	~$6.12
Azure ML (LoRA on 70B, A100)	4 hrs compute	~$14.68

Optimization: Use gpt-4o-mini fine-tuning first (cheapest). Only fine-tune gpt-4o if mini doesn't meet quality bar.

Hyperparameter Search Strategy

# Grid search (small sets)
python scripts/hyperparam_search.py --strategy grid \
  --lora-rank 8,16,32 --lr 5e-5,1e-4,2e-4 --epochs 2,3,5

# Random search (larger sets)
python scripts/hyperparam_search.py --strategy random \
  --trials 10 --budget $50

Production Readiness Checklist

[ ] LoRA rank optimized (start small, increase if underfitting)
[ ] Learning rate validated (no training instability)
[ ] Epochs set to avoid overfitting (val loss not increasing)
[ ] Training data quality reviewed (no duplicates, balanced categories)
[ ] Compute right-sized (don't use A100 for 7B LoRA)
[ ] Cost per training run documented
[ ] MLflow experiment with all hyperparameters logged
[ ] Fine-tuned model compared to base on eval set
[ ] Model registered with version and training metadata

Output: Tuning Report

After tuning, document:

Best hyperparameters found and why
Training cost vs quality trade-off curve
Compute recommendation for this task
Data quantity/quality recommendations for next iteration

tune-fine-tuning-workflow— agent skill

Is tune-fine-tuning-workflow safe to install?

Tune Fine-Tuning Workflow

When to Use

Tuning Dimensions

Dimension 1: LoRA Hyperparameters

Dimension 2: Data Quality Optimization

Dimension 3: Compute Selection

Dimension 4: Azure OpenAI Fine-Tuning Parameters

Dimension 5: Cost Per Training Run

Hyperparameter Search Strategy

Production Readiness Checklist

Output: Tuning Report

Audit the pieces. Scan the whole. Decide.