Pointing at your data
Structure and statistics, never content.
The Dataset control sits beside Templates, because the two answer the same kind of question: what the design starts from, and what it has to fit. Choose a file or a folder, and NEURAX reads how it is shaped. Choosing another one later replaces it once the new one is read; a dataset that cannot be read leaves the one in use unchanged and says why.
| Kind | What it expects |
|---|---|
| Tabular | .csv, .tsv, .parquet — one row per sample. Column types and missing values are read from the header and a scan. |
| Records | .jsonl, .ndjson — one record per line; .json — one record per line, or one array of records. |
| Images | A folder per class. The subfolder names become the labels. |
What comes back is the shape of the data: how many samples, how they are split, how balanced the classes are, and what could not be read. Those figures fill in the fields the design needs — number of classes, channels, image size — that you would otherwise type by hand.
Page 13 of 33. This guide is in English for now. It is the same guide that is built into the NEURAX desktop app (the ? button, or F1). Found a mistake? Tell us.