Datasets

A dataset is a typed-column data table. It is the data you run experiments over and the records your team annotates in labeling jobs. Unlike a spreadsheet, every column has a declared type and an editing widget, so values stay clean and forms render the right control.

Use a dataset when you need structured, reusable input/reference data — a question bank, a set of cases to triage, gold labels to evaluate against, and so on.

Typed columns & widgets

Each column carries a type (how the value is interpreted) and a widget (how you edit it), plus per-column settings.

Type	Use for	Typical widgets	Per-column settings
Number	Numeric values	Input	—
Ordinal	Ranked scales	Select, Radio	Ordinal scale (ordered options)
Text	Free text or categories	Input, Textarea, Select, Radio	Enum options (for Select/Radio)
Boolean	Yes/no flags	Checkbox	—

Available widgets are Input, Textarea, Select, Radio, and Checkbox. Configure columns in the ColumnEditor (add, rename, retype, set enum options or an ordinal scale).

Rows store their values as JSONB data keyed by column name. For backward compatibility, the legacy input / reference fields are kept in sync automatically.

Create a dataset

1. From the Research hub or an investigation’s Datasets tab, create a new dataset and give it a name.

2. Open the ColumnEditor and define your columns — pick a type and widget for each, and set enum options or an ordinal scale where relevant.

3. Add records: use the inline-editable table (RecordsTab) or open the RowForm to fill a record field by field with each column’s widget.

Import CSV or JSON

The ImportModal ingests a file with a dry-run preview, so you confirm the schema before anything is written.

1. Choose a CSV or JSON file to import.

2. Review the dry-run preview. Chatty infers a type for each column; you can edit the schema (types, widgets, column names) before committing.

3. Pick a mode: append to add rows to the existing dataset, or replace to overwrite it.

4. Commit. The records are written using the schema you confirmed.

Caution

Replace discards existing rows. Use append to keep what is already there.

Edit records

• RecordsTab — an inline-editable table; click a cell to edit it with the column’s widget.
• RowForm — a full-record form, useful for longer text fields and careful entry.
• CellRenderer — renders each value according to its column type and widget.

Append from a trial

Experiment outputs can flow back into a dataset. From the experiment results, the AppendFromTrialModal lets you append a trial’s outputs as new rows — handy for building a dataset of model generations or for curating results into gold data. See Experiments & Trials.

Vectors: what was embedded, with what

A dataset can be vectorised: every row of a version becomes a vector with an embedder, and a sample of the rows is laid out on a plane so you can look at it. Open the Vectors section in the dataset panel, tick the columns that make up the text (they are joined as column: value, one per line, in the dataset’s column order), pick the embedder (the workspace’s, the one the platform lends, or any provider model you have a key for), choose a projection, and vectorise.

Three things are worth knowing:

• What gets vectorised is a version. Vectorising the live table pins it first, exactly as running an experiment does, so every vector can be attributed to a row at a digest. Pick a version in the panel to vectorise that one instead.
• The same version, model and columns are one vectorisation. Press the button twice and you get the same row back. Asking for another projection (PCA or t-SNE) of an existing one only redraws the sample; nothing is embedded again. A different model is a second vectorisation, kept side by side.
• The picture is computed once, on the server, over at most 3 000 rows chosen at a fixed stride. The browser only draws the points it is handed, so a large dataset does not weigh on the page.

When the workspace’s embedder is no longer the one a vectorisation was made with, the row is marked other embedder. Retrieval against that vectorisation keeps using its own model (two models never give comparable vectors), so the badge is a prompt to vectorise again, not a broken search.

Shared & global datasets

Datasets can be global/shared: the same dataset can be attached to many investigations (M:N) and copied across workspaces. This lets a question bank or gold set serve several lines of research without duplication.