菜单
Orchestrate the full active learning loop: train, analyze, push to CVAT, wait for review, pull, merge, retrain.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Set up autonomous training monitoring — creates cron jobs to track long-running training, auto-continue pipeline when training completes.
Audit YOLO dataset quality — class distribution, annotation quality, image stats, and improvement suggestions.
Analyze YOLO training runs — compares to baseline/best, checks per-class regression, analyzes training dynamics and tune convergence, writes actionable recommendations.
Run autonomous YOLO training experiments — reads training-plan.md, assesses bottlenecks, acts strategically, and delegates HP optimization to model.tune().
Profile YOLO model inference speed, FPS, and size across image sizes and export formats.
Initialize a new YOLO project — detects your dataset's starting state and routes through the right tools.
| name | active-learning |
| description | Orchestrate the full active learning loop: train, analyze, push to CVAT, wait for review, pull, merge, retrain. |
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.
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
Use the /cvat-push skill:
yolo-cvat push --from-analysis reports/uncertain_images.txt
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.
When the user says annotations are done:
yolo-cvat pull --task <id> --output datasets/corrected_batch_N
yolo-validate datasets/corrected_batch_N
yolo-merge --sources <existing_labels> <corrected_labels> --output <merged_labels>
yolo-validate <merged_dataset>
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.
Delegate training decisions to the refactored /experiment skill.
The reasoning loop will use the updated dataset profile to decide what to try.
/experiment
Compare metrics from before and after the new data:
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]
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
experiments/active_learning_log.md