Missing images, mismatched image tokens, exact train/eval overlap. Inspect the row, field and suggested fix in a real report — without installing anything.
The examples below are generated by the CLI from the repository's tiny synthetic fixtures.
3 train records + 1 eval record. A missing file, an image-token mismatch and cross-split duplication.
Filter by severity or search for a rule such as missing_image. Every finding points to a record and a field.
2 train records + 1 eval record. Local color-square images with matching prompts and no detected overlap.
A passing report covers the supported checks. Label quality and the training pipeline still need separate validation.
These are two independent synthetic datasets, not an automatic repair demonstration. The tool checks data; you decide how to fix it. HTML reports are in English.
This site only serves bundled examples. To inspect private datasets, install the Python CLI and generate reports on your machine.
messages or ShareGPT conversations.vlm-data-doctor check \
examples/clean/train.jsonl \
--eval examples/clean/validation.jsonl \
--strict --format html \
--output clean-report.htmlRun from the repository root and open clean-report.html. Use a new output filename each time; reports are never overwritten.
The auditor checks record structure, image decoding and exact overlap. It does not grade answers, detect cropped or recompressed near duplicates, or predict model performance.
Each published report comes from the installed CLI. The site build verifies both exit codes and all expected finding counts before deployment.