Enterprise Client Checklist

The Dataset Readiness Checklist

Reusable across CS, Sales Enablement, Platform onboarding, and Academy. Print it, share it, and gate every project on it. The questions below come straight from the data collection and annotation and steps of a CV project guides — collapsed into one page that a non-engineer can run.

# Ultralytics — Dataset Readiness Checklist

Project: ____________________________________________________
Owner:   ____________________________________________________
Date:    ____________________________________________________

## 1. What are we detecting?
   Vision task: [ ] detect  [ ] segment  [ ] classify  [ ] pose  [ ] OBB
   Class list:  ___________________________________________
   Class definitions written? [ ] yes  [ ] no

## 2. Where will the model run?
   [ ] Cloud GPU   [ ] Cloud CPU   [ ] Edge (Jetson/Pi)   [ ] Mobile   [ ] Browser
   Latency target: ____ ms   Min hardware: __________________

## 3. What visual variation exists?
   Times of day:    _______________________________________
   Weather:         _______________________________________
   Cameras / sites: _______________________________________
   Operators:       _______________________________________
   Product variants:_______________________________________

## 4. What edge cases matter?
   _______________________________________________________
   _______________________________________________________

## 5. Do we have negative examples?
   [ ] Yes — ___% of dataset is background images (target 0–10%)
   [ ] No  — collect before training

## 6. Are labels consistent?
   [ ] Labeling guide written
   [ ] Calibration round done; inter-rater agreement ___%
   [ ] QC report attached

## 7. Did we check for leakage?
   Split unit: [ ] image  [ ] scene  [ ] camera  [ ] day  [ ] location
   No scene / camera / day appears in more than one split? [ ] yes  [ ] no

## 8. Is the test set truly unseen?
   [ ] Locked. Touched only once for go/no-go.
   [ ] Still being tuned against — STOP, re-split.

## 9. What performance metric defines success?
   Metric: ______________________ Target: ____________
   Acceptance test: _____________________________________

---

Volume targets met?
   [ ] ≥ 1500 images / class
   [ ] ≥ 10,000 labeled instances / class
   [ ] OR documented exception: ________________________

Ready to fine-tune Ultralytics YOLO26?
   [ ] Yes — start with the nano checkpoint matching the task selected in #1
       (detect: `yolo26n.pt`  /  segment: `yolo26n-seg.pt`  /  pose: `yolo26n-pose.pt`
        /  OBB: `yolo26n-obb.pt`  /  classify: `yolo26n-cls.pt`)
   [ ] No  — fix item ___ first

How to use it

At kickoff. Walk through it with the customer or stakeholder. Item 9 (success metric) is usually the one that surfaces real disagreement; spending an hour on it saves the project.

Before collection. Items 1–4 set the dataset spec. Don't capture data without them filled in.

Before annotation. Items 5–6 gate the labeling stage.

Before training. Items 7–8 plus the volume targets gate the first training run.

Before retraining. Re-run the whole thing. New classes, new collection, new annotators — each can break a previously-passing item.

Why a single page

The checklist works because it fits on a single page and uses words a non-engineer can verify. Customer success can run it without engineering present. Sales enablement can use it in pre-sales conversations. Platform onboarding can attach it to a workspace. Docs can link to it. Academy gates the certificate on it.

Same artifact, every team. That's the point.

What this course got you

You can now:

• Translate a business goal into a vision task and class list.
• Write a dataset specification before collecting.
• Plan a representative collection across cameras, times, weather, and operators.
• Annotate consistently with a written guide.
• Run a structured QC pass.
• Split cleanly without leaking.
• Use augmentation that reflects deployment.
• Decide objectively whether the dataset is ready.
• Run the first fine-tune and read its results.
• Hand a customer or your team the checklist.

The next time someone says "we're thinking of a CV project," you have the conversation, the spec, and the checklist on day one — not week six.

Where to go next

• Train your first YOLO model — convert your ready dataset into the YOLO label format, write the data.yaml, and run model.train(...).
• Fine-tuning guide — the deeper reference for layer freezing, two-stage training, and optimizer choice once defaults aren't enough.
• Build with Ultralytics Platform — manage the full lifecycle (dataset → cloud training → endpoint → monitoring → retrain) in one place.
• Ultralytics YOLO in Production — export, serve multi-stream, and observe the model after deployment.