Helix — Training Data Studio

Your guide

Chat with Riu

Riu asks plain-English questions, sets up Helix for you, then starts collecting training data.

Setup 0%

Message Riu

Riu is ready

How Riu works

Asks what you’re training
Learns topics & quality bar
Saves plan, formats & goals
Starts mining (survives logout)

Collected so far

Nothing yet — start the chat.

Shortcuts

Home

How Helix works

You don’t need to know programming. Follow the steps — Helix’s AI helpers do the heavy lifting.

Current plan

Loading your plan…

Needs your attention

Questions for you

When the AI is unsure, it will ask you here.

Activity

Recent helper runs

Step 1

Describe what you’re collecting

Write this like you’re briefing a research assistant. Plain English is perfect.

Draft

Project name A friendly name, e.g. “Customer support AI” or “Sales coach”.

What is this AI about? The topic or business area.

What should Helix collect? Your goal for the training data.

Key questions to answer One per line. Optional.

Topics / categories Comma-separated

Where to look Comma-separated

How many good examples per topic? One topic per line: topic, then a number. Example: billing 40

Example formats to use Names of formats from step 2, comma-separated.

Extra instructions for the AI helpers Optional. Tone, rules, things to avoid.

How the final file should look Optional note for downloads.

Advanced options (optional)

Success goals, output shape

Success goals (one per line)

Formats in this plan

Save formats in step 2, then list their IDs above.

What the plan drives

Targets set here become the per-topic progress on Home, and tell the helpers which topics to weight next.

Tip

Plain English is enough. You can refine later without losing saved data.

Step 2

Your example formats

Each format is a type of training example — “question + answer”, “situation + recommendation”.

Formats Select a row to edit

Create or edit

New format

Friendly name Short ID Letters, numbers, underscores. e.g. support_reply What is this format for? Sample question / input Sample perfect answer / output Why this answer is good (optional) File style when downloading

Advanced: technical template (optional)

Structure template

Step 3 · Mining batches

Run collection batches

Jobs keep running even if you sign out. Come back later — data is saved to your account.

Quality ↔ Cost Mode 2 — High quality

1 · Best quality
All 15 agents 4 · Lowest cost
Ultra lean

Core AI helpers + code for mechanical steps.

Batch size (max 10) How many items each batch processes

Number of batches Auto-runs one after another

Auto-run next batch after each finishes

Estimated total time: calculating…

Optional note

Your jobs

Active and recent mining / synthesis jobs. Safe to leave this page.

Single helper (optional)

Click a card, then “Run one helper only”.

Step 4 · Your account library

Gold & synthesized data

Everything here is stored in your user account forever (no auto-delete). Defaults: 5,000 gold × 4 variations → up to 20,000 synthesized.

Collection goals

How much gold to mine, and how many variations per gold example.

Gold examples goal Vetted real-world / curated rows (default 5000) Variations per gold example Default 4 → 5000 × 4 = 20,000 synthesized

Synthesized goal: 20,000

Automatically save approved examples into my gold library

What to vary when synthesizing

Pick which attributes helpers should change while keeping the answer correct.

Quality ↔ Cost (synthesis) Mode 2

1 · Best (LLM) 4 · Lowest (templates)

Batch size (gold rows, max 10)

Number of batches

Auto-run batches (keeps going after logout)

Estimated total time: —

Gold in your account

Real-world verified / curated examples you own.

Synthesized in your account

Controlled variations of your gold — also stored forever.

Step 5

Download your training data

Export pipeline dataset versions, or your permanent account library (gold + synthesized).

Name this version e.g. March pack or version_1

Collect great examples for your AI

Chat with Riu

How Helix works

Questions for you

Recent helper runs

Describe what you’re collecting

Your example formats

New format

Run collection batches

Your jobs

Single helper (optional)

Gold & synthesized data

Collection goals

What to vary when synthesizing

Gold in your account

Synthesized in your account

Download your training data

Saved versions