LoRA Training Suite: Dataset Doctor + One-Confirm Trainer
Bring a folder of images, leave with a trained LoRA. Catches dataset problems before the GPU runs, then trains after one confirmation.
Free
Works with the AI tools you already use
Claude Code
Cursor
Codex CLI
GitHub Copilot
Gemini CLI+20 moreClaude CodeCursorCodex CLIGitHub CopilotGemini CLI
OpenClaw
Manus
Grok+17 agentsLoRA Training Suite: Dataset Doctor + One-Confirm Trainer
Example session with this skill installed
Check my LoRA dataset at mychar/ before I train. Trigger is mych4r, I'm planning 10 epochs at 1024x1024.
- Read your context and instructions
- Compiled the lora training suite
- Generated the document
`markdown
Demo dataset: 14 images with problems planted on purpose. The report and dry-run blocks are verbatim output from a real run of this skill.
Summary
Not ready to train: FAIL. One image is truncated and will crash training. Three caption problems would quietly weaken the LoRA. I dry-ran five one-line fixes below; nothing has been changed yet.
# Dataset Doctor — ⛔ FAIL
`mychar`
- Images: **14** · effective: **70** · total steps: **700**
- Issues — critical 1, high 3, medium 6, low 0, info 0
Findings
| Severity | Problem | File | Fix |
|---|---|---|---|
| Critical | Cannot be decoded (truncated) | img_14.png | quarantine-corrupt |
| High | No caption file | img_06.png | Tag it (see below) |
| High | Trigger mych4r in only 84.6% of captions | img_08, img_09 | add-trigger |
| High | Multi-line caption: the trainer reads only line 1 | img_07.txt | I merge it into one line |
| Medium | Byte-identical duplicate | img_02.png = img_13.png | dedupe |
| Medium | CMYK image | img_05.jpg | to-rgb |
| Medium | Short side 640 px, target 1024 | img_10.png | Replace the source (no auto-fix) |
| Medium | Source-noise tags watermark, signature | img_11.txt | strip-tags |
| Medium | silver hair, purple eyes in 100% of captions | all | Your call (see below) |
Planned fixes (dry-run, nothing changed)
[move] mychar\5_mych4r\img_14.png -> mychar\_quarantine\img_14.png (OSError: image file is truncated)
[move] mychar\5_mych4r\img_14.txt -> mychar\_quarantine\img_14.txt (caption of img_14.png)
[move] mychar\5_mych4r\img_02.png -> mychar\_quarantine\img_02.png (duplicate of img_13.png)
[move] mychar\5_mych4r\img_02.txt -> mychar\_quarantine\img_02.txt (caption of img_02.png)
[convert] mychar\5_mych4r\img_05.jpg CMYK -> RGB (original backed up to mychar\_quarantine\img_05.jpg)
[rewrite] mychar\5_mych4r\img_08.txt insert trigger 'mych4r' at tag 4
[rewrite] mychar\5_mych4r\img_09.txt insert trigger 'mych4r' at tag 4
[rewrite] mychar\5_mych4r\img_11.txt removed 2 tag(s)
Plus img_07.txt merged to one line: tags first, then . before the sentence.
Needs your decision
- img_06.png has no caption. Tag it with
tag_dataset.py(WD14), or write one by hand. - img_10.png is 640 px. Swap in a 1024 px+ source if you have one.
silver hair / purple eyes are in every caption. Remove them if they should come with mych4r automatically; keep them if you want to change hair or eye colour by prompt.
Reply confirm to apply. Originals go to mychar/_quarantine/; nothing is deleted.
After "confirm" (re-run)
| Metric | Before | After |
|---|---|---|
| Verdict | FAIL (exit 2) | WARN (exit 0) |
| Issues critical / high / medium | 1 / 3 / 6 | 0 / 1 / 2 |
| Trainable images | 14 | 12 |
| Trigger coverage | 84.6% | 100% |
| Total steps (10 epochs) | 700 | 600 |
Next steps
- Caption img_06.png, then re-run the doctor.
- Hand off to
lora-trainer: 600 steps is under its ~1500-step first-run budget, so it will raise repeats/epochs and show the confirmation card before anything trains.
lora-training-suite-dataset-doctor-one-c.pdf
PDF · document
Example file from a real run - the skill writes it into your workspace.
Connects securely to your tools. The creator never sees your data.
About this skill
Most bad LoRAs are decided before training starts: a corrupt image, a caption the trainer silently cuts off, a trigger word missing from a fifth of the captions, a step budget that is far too low or too high. The trainer only checks that the folder exists and has images. This suite checks the rest, fixes what can be fixed safely, and then runs training with one plain-language confirmation.
Three skills, three jobs
dataset-doctor - Audits a kohya-style dataset and returns PASS, WARN, or FAIL with fixes in priority order. It checks image count and step budget, resolution and aspect buckets, exact and near duplicates, corrupt and non-RGB images, missing, empty, or multi-line captions, trigger coverage, and tag hygiene (watermark/signature tags, tags present in every caption). Offline, takes seconds, no GPU. Works before any kohya-based trainer, not only SD-Trainer.
lora-trainer - Give it a folder of images. It organizes the folder, auto-captions, runs the doctor as a gate, picks repeats and epochs from your image count and detected VRAM, and shows one confirmation card. After you reply "confirm" it launches lora-scripts-next (SD-Trainer) through its local API and watches the log. An expert path accepts your own presets, dims, and learning rates.
lora-pipeline - Give it just a character or style name. It collects images from Danbooru, curates and captions them, runs the doctor, trains through lora-trainer, renders a fixed-seed sample gallery in ComfyUI, and fills the Civitai upload form up to Draft. You review and click Publish yourself.
Safety model
- Every repair prints its plan first and changes nothing until you confirm.
- Displaced files move to _quarantine/ inside your dataset. Nothing is ever deleted.
- Training never starts without the confirmation card.
- Civitai uploads stop at Draft. Nothing is published automatically.
Models
Anima first, with SD1.5, SDXL, and Flux through the same trainer. Caption rules were checked against the Anima model card and the trainer's own source code, not forum folklore (for example: kohya reads only the first line of a .txt caption, so the doctor flags multi-line captions).
Open source
MIT licensed, same code as github.com/Rinne414/lora-training-skill.
How to install
Works the same in every agent - Claude, Cursor, Codex, Copilot and 20+ more.
- 1
Download the ZIP
Free skills download straight away. Paid skills unlock right after purchase.
- 2
Unzip into your skills folder
Every agent reads skills from one folder on your machine. Drop the unzipped folder in there.
- 3
Ask your agent to use it
Restart the agent if it was already running. It picks the skill up automatically - no config needed.
Skills folder by agent
Click the path to copy it. Create the folder if it does not exist yet.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill passes our trust checks below.
Security scanned
Passed our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scanned
Verified clean 1 day ago
- Free to download with an account