Active Learning Loop

Orchestrate the complete cycle of model improvement through human feedback: train -> analyze -> push uncertain images to CVAT -> wait for human review -> pull corrections -> merge -> retrain.

Workflow

1. Check Current State

Determine where we are in the loop:

Has baseline run?
├── No → Run: /experiment baseline
└── Yes → Continue

Has analysis been run?
├── No → Run: /analyze
└── Yes → Continue

Has uncertain_images.txt?
├── No → Run: yolo-analyze --model <best> --dataset <ds>
└── Yes → Continue

2. Push Uncertain Images to CVAT

Use the /cvat-push skill:

yolo-cvat push --from-analysis reports/uncertain_images.txt

3. Wait for Human Review

Tell the user:

Images have been pushed to CVAT for review.
Please annotate/correct the images in CVAT, then tell me when you're done.

CVAT tasks: [list task IDs and URLs]
Images to review: [count]
Priority: [false negatives first, then uncertain]

Do NOT proceed until the user confirms annotations are complete.

4. Pull Corrected Annotations

When the user says annotations are done:

yolo-cvat pull --task <id> --output datasets/corrected_batch_N
yolo-validate datasets/corrected_batch_N

5. Merge with Existing Dataset

yolo-merge --sources <existing_labels> <corrected_labels> --output <merged_labels>
yolo-validate <merged_dataset>

5b. Re-Profile After Merge

Run /review-dataset to update the dataset profile in training-plan.md. The profile may have changed (new class distribution, different object sizes). The experiment reasoning loop needs current data.

6. Retrain

Delegate training decisions to the refactored /experiment skill. The reasoning loop will use the updated dataset profile to decide what to try.

/experiment

7. Compare Before/After

Compare metrics from before and after the new data:

• Overall mAP50-95 change
• Per-class AP changes
• Specifically check classes that had false negatives

8. Write Active Learning Report

Create experiments/active_learning_log.md:

## Active Learning Iteration N
- Date: YYYY-MM-DD
- Images added/corrected: X
- mAP50-95 before: X.XXXX
- mAP50-95 after: X.XXXX
- Delta: +/- X.XXXX
- Classes most improved: [list]
- Next action: [recommendation]

Decision Tree

First iteration ever?
├── Yes → Establish baseline first, then analyze
└── No → Check if previous iteration improved metrics

Metrics improved after new data?
├── Yes → Deploy updated model to CVAT (/cvat-deploy)
│         Continue to next iteration
└── No → Investigate:
         ├── Data quality issue? → Review annotations more carefully
         ├── Overfitting to corrections? → Increase augmentation
         └── Plateaued? → Suggest dataset expansion or architecture change

More than 3 iterations with <0.5% improvement?
├── Yes → Suggest stopping: diminishing returns
└── No → Continue loop

Guardrails

• NEVER skip the merge validation step — corrupted merges ruin models
• ALWAYS compare metrics before and after new data is added
• Document every iteration in experiments/active_learning_log.md
• If metrics regress after adding data, STOP and investigate before continuing
• Maximum recommended batch: 200 images per iteration (annotator fatigue)
• Wait for user confirmation at every human-in-the-loop step