Creating a dataset in Data Craft

Every training run needs a dataset. Data Craft is the Ertas tab where datasets live: it shows each dataset's row count, file size, schema, and a preview, and it is where you build new ones. There are two ways to start a dataset and two ways to fill one you created.

Start a dataset two ways

Data Craft gives you two top-level entry points, side by side:

• Create a dataset builds an empty dataset that you then fill.
• Upload a file builds a dataset directly from a JSONL file you already have.

They are peers. Uploading a file is not a step inside creating one. It is a separate door for when your data already exists.

Create a dataset

Creating a dataset asks two questions: a name and a row format. That is it. The row format is the JSONL schema every row will follow (text, instruction, input, conversations, or messages), and it locks after creation because every later step depends on it. Your objective and the conversational roles come later, in Prompt Studio, when you are ready to generate.

Upload a file

If you already have data, upload it directly. Data Craft accepts a .jsonl file from your disk (up to 1GB) or a HuggingFace dataset URL (tick the third-party review box to confirm you have the right to use it). Ertas reads the file, detects the per-line format, and builds the dataset from it. The supported per-line formats are the same five schemas: text, instruction, input, conversations, and messages.

Fill a created dataset

A dataset you created starts empty. There are two ways to add rows: write them by hand, or generate them with your own AI tool through Prompt Studio.

Write rows by hand

For a small, high-quality set, type rows in directly. Choose Add row, write the user turn, then the assistant answer, and save. New rows start in a not accepted state, and only accepted rows ever train.

Hand-writing is slow but exact. It is the right tool for seed examples, a handful of tricky edge cases, or a golden set you want under full control.

Generate rows with your own AI tool

To scale up, Ertas does not run generation itself. Instead, Prompt Studio turns your context into a prompt, you run that prompt in your own AI tool (ChatGPT, Claude, or anything else), and you paste the results back. Because the generation runs in your tool, it does not consume Ertas credits.

Set the context first

Step 1

Open Prompt Studio and define your objective (one or two sentences on what the model should do) and the conversational roles: who the user is and who the assistant is. Setting the context first is the whole game. Data quality is the single biggest lever on the quality of the model you train, and the context is what makes the generated rows look like your real distribution instead of generic filler.

The modal says it plainly: "Generate via your AI tool. Define your objective on the left to unlock prompt generation, or paste JSONL directly on the right."

Copy the prompt Ertas builds

Step 2

From that context, Ertas drafts a generation prompt. It is fully editable, so tighten it however you like, and the number of rows to generate is set inside the prompt itself. When it reads the way you want, copy it.

Run it in your AI tool

Step 3

Paste the prompt into ChatGPT, Claude, or whichever model you prefer, and run it. It returns JSONL rows in your chosen format.

Paste the rows back into Prompt Studio

Step 4

Copy the JSONL the model produced and paste it as text into Prompt Studio's import field. Ertas parses it and adds the rows to your dataset as drafts, labelled AI DRAFT. The success modal confirms it: "Imported N rows. They're sitting in your dataset as drafts. Review and accept them in the table."

Review and accept

Step 5

Drafts do not train. Read the imported rows in the table and accept the ones that are good. Only accepted rows are sent to the trainer when you start a run. This review pass is where a generated batch becomes a genuinely good dataset, so do not skip it. For what to look for, see Dataset quality. For the techniques that make generated rows better in the first place, see Dataset synthesis.

What's next