WIDECHAOS / SPLIT BOUNDARY AUDIT 0.2

Keep groups together.

Plan train, validation and test splits while keeping declared groups and exact duplicate text together.

See the achieved ratios, inspect label coverage, and export every record. Targets are approximate; this greedy planner does not promise the best possible balance.

1. Describe the indivisible groups

Each line: id, group, label, text. Up to 5,000 records and 2 MiB UTF-8. The bundled 36-record example is synthetic.

What should a group mean?

A caller-defined indivisible evaluation unit, such as a document or a template family. A class label or population cohort usually does not describe that unit. Missing groups are rejected; the planner cannot infer trustworthy groups for you.

Exact text duplicates connect groups into larger allocation units. NFC/whitespace matches only produce review warnings; they do not connect groups. No rows are discarded.

Native JSON.parse keeps the last duplicate object key. This application does not detect duplicate keys within a line.

2. Set approximate targets

Row percentages · positive integers, total 100

Loading synthetic example…

Input and settings stay in your browser. Editing them clears the previous plan and download links.

3. Review the plan

Plan the sample or your own grouped data to inspect its actual distribution.