Prepare mode

Prepare mode is the other half of MLabel: not labeling, but getting work to labelers and collecting it back.

Reach it from the Mode menu, or ⌘/Ctrl+⇧+P.

⌘/Ctrl+⇧+L goes back to labeling. A config must be loaded first — Prepare validates files against it.

Drop first

You do not pick an operation and then find files. You drop files, and MLabel works out what you are trying to do:

What you drop	What it proposes
One file matching the input schema	Split
Several *-output files	Join outputs
Several *-remaining files	Join remaining

It always asks before acting. The proposal is a guess from filenames and contents, and joining the wrong set of files is not something to discover afterwards.

Splitting

Choose how many parts. MLabel divides the records into contiguous chunks whose sizes differ by at most one, with the larger chunks first — 10 records into 3 parts gives 4, 3, 3.

Files are written next to the source, named predictably:

data.csv  ──▶  data-part1-of-3.csv
               data-part2-of-3.csv
               data-part3-of-3.csv

The preview shows exactly which names will be written before you commit. Existing files are never overwritten — a collision fails the operation and names the files.

Each part is a valid input file for the same config, so each labeler just opens theirs.

Joining

When labelers send work back, you have a pile of *-output files (and possibly *-remaining ones). Drop them all in.

Every file is checked before anything is written:

• Headers must match exactly, including order — cells are positional, so a reordered header would silently shift every value. A mismatch is a blocking error.
• Output files are validated against the output schema: every value must coerce to its declared type, required fields must be present, and provided values must pass their rules. Any failure blocks the join.
• Remaining files are validated against the input schema, but non-blockingly. Labeling itself tolerates a cell that will not coerce, so joining does too — you get a warning.
• Duplicate rows are flagged as warnings. This is what catches the same part being submitted twice.

You choose where the result is saved. The default name strips the part and kind suffixes and adds the join kind: data-part1-of-3-output.csv → data-output-joined.csv.

The round-trip caveat

A joined *-remaining file is value-faithful but not byte-faithful — see Exporting. It reloads into MLabel perfectly, which is what the workflow depends on.

Prepare holds no state

Files are re-read from disk when you run an operation, not cached from when you dropped them. Editing a file after analysing it and before running is safe: you get the current contents, not a stale snapshot.

For the full workflow of running a project this way, see Distributing work.