MIRA / DATASET LAB

Mira Dataset Studio

Secure session
BUILD BETTER MATCHING DATA

Shape the signal behind Mira.

Create clean language pairs, generate balanced combinations, and label them with as little friction as possible.

—offers
—wants
—to label
01 · BUILD THE BANK

Turn an activity into useful language.

Draft an offer and a want from a plain-language activity, then tune the wording before it enters the bank.

Fast drafting · 9 variants each
01Drafting desk

Describe an activity

Use a concrete skill, class, service, or activity. The fastest results come from short, natural phrases.

02Live item bank

What Mira can draw from

Fresh items make better combinations. Keep both sides growing.

Offers 0

    Wants 0

      Back up the item bank

      Save every offer and want to a file on your computer, then restore it later — after a redeploy, an update, or on another device — without retyping anything.

      TEAM IDENTITY

      Choose who is working.

      Pick your name, enter your private code, and Mira will keep your contributions and mission progress attached to you.

      No user signed in
      TEAM MEMBERS 0 users
      No users yet. Add the first team member to get started.

      BULLETIN BOARD

      Who's moving the dataset?

      Add users to start the board.
      CREATE MISSIONReady

      Write 25 new pairs.

      Set a target, start the mission, and watch your unique offer + want pairs add up as you save them.

      Flag
      0 / 25 pairs written25 to go
      01Writeshape the language
      02Pairsave both sides
      03Goalcross the finish line
      SAVED PRESETS
      No presets yet — save a goal you like.
      COMPLETED MISSIONS
      No completed missions yet.

      DATA QUALITY

      Pair quality, right now.

      Re-examine only the offer + want pairs currently in the bank. Label history, save attempts, and old snapshots never affect this check.

      Current pairs0offer + want pairs in the bank
      Average pair length0 wordsoffer + want together
      CURRENT CHECK

      What is in the bank now

      not examined yet
      Re-examine the current pairs to see the latest assessment.
      TOP THEMES

      What the dataset talks about

      OFFERSwhat people can give
      No pairs yet.
      WANTSwhat people need
      No pairs yet.
      02 · COMPOSE THE QUEUE

      Generate a balanced set of combinations.

      Mira rotates through varied 1×1 → 3×3 size buckets and follows your chosen Yes/No balance using the source-phrase provenance already attached to your item bank.

      40/60–50/50 range
      03Combination engine

      How much signal do you need?

      Start with a focused batch while you are iterating. You can always generate more later.

      04Live balance

      Shape of this run

      Offers × wants, with varied sizes from 1×1 through 3×3 and your selected Yes/No target shown after the run.

      0new
      03 · HUMAN JUDGMENT

      Make the match decision fast.

      Read the offer and want language side by side. Shared wording is highlighted automatically to speed up review.

      Y Match or N No match
      0 unlabeled · 0 total
      LABEL HISTORY

      Review activity

      Select label history to release, export, or remove. History and release membership stay linked: removing a label record returns its activity to the unlabeled queue so it cannot appear in the labeled export.

      0 selected
      Loading history…
      04 · SHIP THE DATA

      Export the labeled dataset when the batch is ready.

      Download one JSON object per line used by the training pipeline, with labels ordered by their saved timestamps.

      JSON · sequential IDs
      05Release package

      Dataset snapshot

      This release is linked to current label history. Only activities that still have an active saved label are included.

      ✓
      Blind labeling preserved

      Exports retain the stored decision and labeler metadata without exposing a suggested answer during review.

      PRE-EXPORT QUALITY GATE

      Check pair reuse before release

      Mira scans the labeled queue for the same normalized offer × want relationship appearing too often. The clean export keeps the earliest appearances and quarantines later tasks containing overused relationships. Database records are never deleted by this check.

      ◎
      No quality check run yet.

      Deleted or unlinked label activities are automatically excluded from this release.

      EXPORT EDITOR

      Edit the release clearly

      Load the current labeled JSONL, make corrections, validate it, then download your edited file. This does not change the database.

      No export loaded.
      06Format preview
      dataset.json
      {"id":0,"offers":["…"],"wants":["…"],"human_label":"yes","labeler":"…","labeled_blind":true,"labeled_at":"2026-08-19T18:44:13.269-07:00"}
{"id":1,"offers":["…"],"wants":["…"],"human_label":"no","labeler":"…","labeled_blind":true,"labeled_at":"2026-08-19T18:45:24.608-07:00"}