LoRA Dataset Quality Checklist
Use this LoRA dataset quality checklist after curation, cleanup, deduplication, resizing decisions, and captioning are finished. It is a release gate for the exact files you plan to train, not another editing pass. The dataset should fail preflight if an image cannot decode, a caption is missing, a trigger token changes spelling, the folder loader finds unexpected files, or important concept coverage exists only in your memory.
Freeze the candidate directory, generate an inventory, and record pass or fail evidence for each section. Fix failures in the appropriate working folder, then rebuild and rerun the entire preflight. Do not patch final files casually after approval. The checklist covers concept definition, visual accuracy, diversity, duplicates, dimensions, metadata, rights, loader behavior, and validation readiness. It ends with a simple go or no-go decision that can be repeated for later dataset versions.
Quick Answer
A LoRA dataset is ready only when all critical gates pass:
- One written concept goal.
- Accurate and readable images.
- Useful variation without dominant accidental correlations.
- No unresolved exact or near-duplicates.
- Model-appropriate dimensions and safe crops.
- Complete, consistent captions or metadata.
- A clean trainer-readable folder.
- Documented rights and provenance.
- Fixed validation prompts and checkpoint plan.
- A frozen manifest identifying the exact files.
Treat any missing image-caption pair, unreadable file, wrong concept example, unclear permission, or failed trainer preview as a stop condition.
How to Use This Checklist
Run the checklist on the final dataset path, not on a general collection folder. Assign every check one result:
- Pass: Evidence confirms the requirement.
- Fail: The dataset must be corrected.
- Not applicable: The requirement genuinely does not apply, with a reason.
Avoid "probably fine." Record counts, filenames, hashes, screenshots, or configuration values. A repeatable preflight is more useful than one person's confidence.
If a file changes after approval, create a new dataset version or invalidate the old report. Pixel edits, caption edits, renames, and configuration changes can all alter training behavior.
Gate 1: Concept Contract
Confirm the project has one concise statement:
Teach [target concept] while keeping [incidental details] controllable.
Pass when:
- Every image supports the same target.
- The target is a subject, object, garment, style, or other defined concept.
- Incidental details are named.
- The target base model family is identified.
- The expected trigger behavior is documented.
Fail when:
- The set mixes unrelated people, products, or styles without a multi-concept plan.
- Team members give different answers about what should be learned.
- The goal is only "make good images."
This gate controls every later judgment. Without it, there is no objective meaning for coverage or contamination.
Gate 2: File Inventory and Integrity
Generate an inventory directly from the final directory.
Record for each image:
- Relative path.
- Stable filename.
- File format.
- Width and height.
- Color mode.
- File size.
- SHA-256 hash.
- Caption or metadata reference.
Pass when every intended image decodes completely, has nonzero dimensions, and appears exactly once in the manifest.
Fail on:
- Corrupt or partially decoded files.
- Unsupported formats.
- Broken alpha channels.
- Case-only filename collisions.
- Cloud placeholders without local content.
- Unexpected hidden or temporary files.
- Manifest entries that do not resolve.
Open a sample through the same library or trainer path used for preprocessing. A file that opens in one desktop viewer can still fail a specific loader.
Gate 3: Visual Accuracy
Review a fresh contact sheet, then inspect every image at 100 percent.
Pass when:
- Identity, object geometry, garment construction, or style is correct.
- Important details are sharp enough to learn.
- Anatomy and generated structures are coherent.
- No watermark, timestamp, interface element, or accidental border contaminates the concept.
- Crops are deliberate and preserve defining features.
- Compression and processing artifacts are minor enough not to become evidence.
Fail if any image contradicts the target, even if it is attractive. A clean image of the wrong product version or unstable generated identity is a concept error.
Record rejected filenames and reasons outside the final training root.
Gate 4: Coverage and Balance
Group the contact sheet by features that should remain controllable.
For a person or character, count:
- Close, medium, and full-body framing.
- Front, three-quarter, and side views.
- Clothing, expressions, lighting, and backgrounds.
- Occluded versus unobstructed defining features.
For an object or garment, count:
- Front, rear, side, and detail views.
- Scale, context, lighting, material, and configuration.
- Views that explain construction and silhouette.
For a style, count:
- Subject categories.
- Composition types.
- Color palettes.
- Scenes, viewpoints, and aspect ratios.
Pass when the target remains stable while incidental variables change enough to prevent one accidental pattern from dominating.
Fail when the dataset relies almost entirely on one outfit, room, pose, subject category, or generated composition. Captions can label repeated details, but captions do not create missing visual coverage.
Gate 5: Duplicate and Weighting Audit
Confirm the frozen set has completed:
- SHA-256 grouping for exact copies.
- Perceptual or visual grouping for resized and recompressed copies.
- Review of crops, mirrors, and source derivatives.
- Burst-frame review.
- Contact-sheet review for semantic repetition.
Pass when each related frame adds distinct evidence and redundant files are quarantined outside the scan root.
Fail when:
- Duplicate groups lack a decision.
- Several files come from one static frame without a documented reason.
- The team removed duplicates but did not update image counts or exposure settings.
Record the scan method and thresholds. Automated similarity is a candidate generator, so final decisions should include human review.
Gate 6: Resolution and Crop Safety
Match the target resolution to the base model and trainer. Confirm aspect-ratio behavior from the actual configuration.
Pass when:
- Sources contain real detail at the selected scale.
- Images are not stretched into squares.
- Bucketing is enabled when mixed aspect ratios require it.
- Tight images retain defining features after resize and crop.
- Minimum and maximum bucket settings are supported.
- Representative portrait, landscape, and square files preview correctly.
Fail when a loader accepts dimensions but the concept becomes too small, distorted, or cropped. Pixel count alone does not pass this gate.
Document any upscaled exception and review it for invented texture.
Gate 7: Caption and Trigger Audit
Captions must match the chosen training method.
Pass when:
- Every image has exactly one expected caption or metadata record.
- No caption is unintentionally blank.
- Trigger spelling and capitalization are consistent.
- Captions describe visible incidental details that should remain controllable.
- Captions do not claim objects, poses, colors, or features that are absent.
- The selected format is consistent.
- Caption shuffle and keep-token behavior are documented.
- Text is valid UTF-8.
Read captions beside thumbnails. A grammatically valid caption can still belong to the wrong file after a rename.
Fail on orphan sidecars, orphan metadata rows, duplicate mappings, stale filenames, or conflicting trigger tokens.
Gate 8: Folder and Loader Contract
Verify the dataset through the exact trainer, script, or loader version planned for the run.
Pass when:
- The configured root is correct.
- Recursive scanning behavior is understood.
- Only intended files are discovered.
- Subsets and repeat values match the written plan.
- Regularization data is separate and correctly mapped when used.
- Validation assets, rejects, originals, and outputs are outside the scan root.
- Dataset preview or caching completes without warnings that alter the sample count.
Compare the trainer's reported image or sample count with your independent inventory. Explain any difference before training.
Fail if a tutorial's folder naming convention is being assumed without confirming that the chosen loader parses it.
Gate 9: Rights, Privacy, and Provenance
Pass when every source has a recorded basis for use, such as ownership, direct permission, or a compatible license.
Confirm:
- Source URL or origin is recorded when relevant.
- License terms and attribution duties are stored.
- Consent and privacy concerns were reviewed.
- Private addresses, documents, screens, and metadata are removed when unnecessary.
- Bystanders and sensitive subjects have been handled appropriately.
- Rights apply to the intended training and output use, not merely viewing.
Fail on unknown provenance or "found online." Public access does not establish permission.
This is a critical gate. Do not use technical cleanup to conceal or bypass a rights problem.
Gate 10: Exposure Plan
Image count is only one part of training exposure. Record:
- Number of images per subset.
- Repeats.
- Epochs or maximum steps.
- Batch size and gradient accumulation where relevant.
- Caption shuffle policy.
- Save interval.
Calculate expected exposure using the trainer's documented behavior. Confirm that one small subset is not unintentionally multiplied far beyond the rest.
Pass when settings reflect the final deduplicated count. Fail when they were copied from a larger or earlier dataset without recalculation.
Repeats cannot repair poor coverage. If one necessary angle is missing, collect it rather than showing the same frame more often.
Gate 11: Validation Readiness
Prepare fixed prompts and seeds before the full run.
Include:
- A neutral concept recognition prompt.
- An unseen background or context.
- A changed incidental detail such as clothing, color, or material.
- A difficult angle or framing request.
- A prompt without the trigger when that test is meaningful.
Record generation settings and the base checkpoint. Decide which training checkpoints will be saved and compared.
Pass when the test set can reveal likeness, prompt obedience, background leakage, composition collapse, and overfitting. Fail when evaluation depends only on the final checkpoint and one favorable prompt.
Validation images and prompts must remain outside the training scan root.
Gate 12: Reproducibility Freeze
Build a release record containing:
- Dataset version.
- Manifest with hashes.
- Trainer and version.
- Base model identifier and revision where available.
- Dataset configuration.
- Caption settings.
- Preprocessing policy.
- Duplicate scan method.
- Validation prompt set.
- Approval date and reviewer.
Make the approved dataset read-only or store it in versioned archival storage. The exact run should be reproducible from this package without relying on an untracked desktop folder.
Pass when changing a file necessarily creates a new manifest result. Fail when edits can occur without detection.
Critical Stop Conditions
Do not begin a full run if any of these remain:
- Wrong or inconsistent concept examples.
- Corrupt images.
- Missing or mismatched captions.
- Unresolved duplicate groups.
- Unclear usage rights.
- Unexpected trainer sample count.
- Destructive automatic crops.
- Trigger inconsistency.
- No fixed validation plan.
A short loader test can diagnose these issues, but it does not waive them.
Final Go or No-Go Review
Mark go only when all critical gates pass and every noncritical exception has an owner, reason, and expected effect.
The approval note should answer:
- What exact concept is being trained?
- How many unique images will load?
- Which caption contract is active?
- What target resolution and bucket policy are active?
- What rights basis covers the files?
- Which prompts will compare checkpoints?
- Which manifest identifies this version?
If any answer requires guessing, mark no-go, repair the source stage, rebuild the final directory, and rerun all gates.
Bottom Line
Dataset quality is not one aesthetic score. It is a chain of verified properties covering concept truth, visual evidence, weighting, metadata, loader behavior, rights, and evaluation.
Freeze only a dataset that can pass the same checklist twice. That discipline makes a training failure diagnosable and a successful run reproducible.
What to Do Next
Freeze the candidate directory and generate a hashed manifest.
Follow lora training dataset guide for the detailed workflow and checks.
Record pass or fail evidence for every critical gate.
Follow best image size for lora training for the detailed workflow and checks.
Run the trainer preview and compare its reported sample count.
Follow lora dataset folder structure for the detailed workflow and checks.