review-dataset
Skill8 starsChanged 8 months agoMITView on GitHub
Audit YOLO dataset quality — class distribution, annotation quality, image stats, and improvement suggestions.
In this repository15 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- active-learningSkill
- analyze-resultsSkill
- annotateSkill
- auto-experimentSkill
- benchmarkSkill
- compare-modelsSkill
- cvat-deploySkill
- cvat-pullSkill
- cvat-pushSkill
- explain-resultsSkill
- monitor-trainingSkill
- review-annotationsSkill
- setup-projectSkill
- training-loopSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 3 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 14
Read yolo-project.yaml for dataset path, class names, and imgsz.
- Distinctive: line 15
If no config, ask user for dataset path.
- Distinctive: line 42
Read the annotated image with Claude vision
- Distinctive: line 50
Incomplete annotations: If a significant % of single-class images have visible objects of the missing class (e.g., body annotated but …
- Very specific: line 63
Run the profiling script with imgsz from yolo-project.yaml (default 640):
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Review Dataset
Dataset quality audit + profiling for architecture selection.
Workflow
1. Load Config
Read yolo-project.yaml for dataset path, class names, and imgsz.
If no config, ask user for dataset path.
2. Run Structural Validation
yolo-validate <dataset_path> --strict3. Deep Analysis
Scan label files for:
- Class Distribution: Count per class, imbalance ratio. CRITICAL if any class <500, WARNING if <1000.
- Annotation Quality: Tiny boxes (<0.02 normalized), edge-clipped, heavily overlapping (IoU >0.9)
- Empty images: Count and percentage
- Split Balance: Compare class ratios between train and val
3.5 Visual Spot-Check (Claude Vision)
Sample random images and visually verify annotation quality using Claude's multimodal vision.
Categorize images by annotation type:
- Both classes: images containing all project classes
- Single-class only: images with only one class (e.g., body without head)
- Empty: negative samples with no annotations
For each category, sample 2-4 images:
- Draw annotations using
python scripts/draw_annotations.py - Read the annotated image with Claude vision
- Assess:
- Are visible objects properly annotated?
- Are there missing annotations (visible objects with no box)?
- Are boxes accurately placed?
- Are class labels correct?
Flag issues:
- Incomplete annotations: If a significant % of single-class images have visible objects of the missing class (e.g., body annotated but visible head not annotated), flag as "incomplete annotations" with estimated scope
- Misclassified: Wrong class assigned to a box
- Phantom boxes: Boxes on non-objects (UI elements, background)
Report findings:
Include visual spot-check results in the audit report with:
- Number of images sampled per category
- Issues found with example filenames
- Estimated % of dataset affected
- Recommendation: auto-label, manual review, or acceptable as-is
4. Dataset Profile for Architecture Selection
Run the profiling script with imgsz from yolo-project.yaml (default 640):
python scripts/profile_dataset.py \
--labels <dataset>/labels/train \
--images <dataset>/images/train \
--imgsz <imgsz> \
--class-names "<comma-separated from data.yaml>"The script outputs structured YAML with:
- Object scale distribution (% small/medium/large at training resolution)
- Min object size (px) at training resolution
- Class-wise profiles sorted by avg object size ascending (smallest first)
- Train/val scale divergence check (flags >15% difference in small_pct)
- Suggested starting point for architecture (head config + scale + reasoning)
Save the output to experiments/dataset_profile.yaml.
5. Write to training-plan.md
Fill the Dataset Summary section in training-plan.md with profile data:
- Total images: train/val counts
- Classes: N — [list]
- Class balance: most/least represented with counts
- Scale distribution: % small/medium/large at imgsz
- Min object size at training resolution
- Avg objects per image
- Train/val divergence flag
6. Write Report
Create experiments/dataset_audit.md with:
- Validation findings
- Class distribution table
- Quality issues
- Dataset profile (full YAML output)
- Architecture suggestion (labeled "Suggested starting point", NOT "Recommended")
- Top 3 recommendations
7. Print Summary
Key findings, profile highlights, and architecture suggestion.
Guidelines
- Be specific: "Collect 400+ bird images" not "collect more data"
- Profile uses training-resolution-adjusted sizes, not native pixel sizes
- Architecture suggestion provides data for agent reasoning — it's not a directive
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

