MLToolslabel · build · tune · ship

Datasets

Dataset Statistics

Before training, look at the data. Drop YOLO, COCO or VOC labels (and optionally images) and get class frequency, boxes per image, box size and aspect-ratio distributions, and image resolution counts. The charts are SVG and can be saved for a report.

Loading the tool…

How it works

01

Drop the dataset

Labels alone give class and box stats; add images for resolution and normalised-size histograms.

02

Read the charts

Class imbalance, tiny boxes and extreme aspect ratios are the three things that most often explain a bad run.

03

Export

Save any chart as SVG or PNG, or the whole summary as JSON or Markdown.

Formats, limits and the numbers that matter

Small boxes

Boxes under 32 px on the training resolution are 'small' in COCO terms and need a higher input size or an anchor-free model to detect well.

Imbalance

A ratio above 10:1 between the most and least frequent class usually calls for oversampling or class weights; the tool prints the ratio.

Resolution

If images vary widely in size, decide the training size from the box-size histogram at that size, not from the image sizes.

Common questions

Does it handle classification folders?

Yes: class folders or a labels CSV give per-class counts and image sizes.

Can I compare train and val?

Drop both split folders and the charts overlay the two.

Text datasets?

JSONL or CSV with a label column gives label counts and text length distribution.

Next in the workflow

Dataset ValidatorFind missing labels, empty files, bad class ids, out-of-bounds boxes and unreadable images before a training run fails.Train / Validation / Test SplitterSplit images with labels, class folders, or CSV/JSONL rows into train, val and test with a fixed seed and stratification.Label Format ConverterConvert detection labels between YOLO, COCO, Pascal VOC and CSV without losing a box.