Orinth
GitHubDownload

Data Recipes

Data Recipes turn source documents you already have into reviewable training records. Add files, choose what the recipe should produce, inspect every generated row, then commit the approved records into Dataset Studio — ready to split, version, and train on.

Supported sources include PDF, DOCX, TXT, Markdown, CSV, and JSONL, up to 20 MB per file. Text is extracted locally when a source is added. Image-only or scanned PDFs are excluded with a warning because recipes do not run OCR.

The two modes

  • Deterministic (default) — chunking and extraction run with no model in the loop. The same inputs produce the same output every time, and nothing is sent anywhere.
  • LLM-assisted (opt-in) — when you supply a key under Settings, a recipe can use a language model to draft instruction/response pairs from your source text. This is off unless you turn it on.

The distinction matters for the local-first promise: a recipe never reaches the network unless you explicitly configure it to.

Building a recipe

  1. Create a recipe inside an LLM-enabled project and name it.
  2. Start from a template or a blank instruction recipe.
  3. Add source documents and resolve any extraction warnings.
  4. Choose the output shape — instruction pairs, question/answer records, or conversational records.
  5. Configure chunk size, overlap, records per chunk, and the generation mode.
  6. Generate and review the records. Edit, approve, or exclude individual rows.
  7. Commit the approved records into the project as a dataset.

The workspace keeps these four stages visible as Sources → Generate → Review → Commit. You can move backward without losing the current preview.

Generation controls

  • Mode — Auto, deterministic Rules, or explicitly LLM-assisted.
  • Prompt style — question and answer, instruction, or conversation.
  • Chunk size and overlap — control how source text is divided while preserving context across boundaries.
  • Records per chunk — controls output density without changing the source material.
  • Model override — optional in LLM-assisted mode; otherwise the configured platform default is used.

Auto mode prefers the configured LLM when a key is available and falls back to deterministic rules when it is not. Choosing Rules guarantees that generation remains local and repeatable.

Review before commit

Generated rows are never silently added to a dataset. The Review stage shows the instruction, optional input, response, and source trace for each record. Exclude weak rows, edit wording, and approve the final set before committing.

From recipe to run

A generated dataset behaves like any other: it lands in Dataset Studio, goes through splits and preprocessing, gets frozen as a version, and feeds a training run. Recipes are a starting point for the pipeline, not a side channel around it.

Provenance

Generated datasets record which recipe produced them, the generation settings, and the source documents behind each record. A model trained on generated data can therefore be traced from its dataset version back to the documents that shaped it.

Try it on the landing page

The interactive product demo includes a static Data Recipe workspace. It reproduces source upload states, generation controls, review cards, and the final Dataset Studio handoff without uploading files or calling a model.