Try it: training-data studio

Build the exact training file a fine-tune reads: write question-and-answer pairs, watch them formatted into the JSONL Unsloth expects, and run the PII scan before anything trains.

All starter content on this page is fictional, written for teaching. It describes a made-up association and no real training run.

1. Pick a teaching use case

Each use case loads two fictional starter pairs, labeled as teaching content. Replace them with your own.

2. Write your pairs

0 pairs (aim for 50 to 200 reviewed pairs in the real file; this studio teaches the format)

3. Watch the JSONL

4. Scan for PII

The scan catches patterns, not names. A human still reads every pair before it trains.

Before and after

The comparison below illustrates the concept. It is not output from a real trained model.

Base model answer (illustration)

Question: When does dockmaster certification expire?

Certification timelines vary by organization. Check with your certifying body for its renewal requirements and deadlines.

After fine-tuning on your Q&As (illustration)

Question: When does dockmaster certification expire?

Dockmaster certification at the Harborlight Marina Association renews every year. Re-certify by finishing the annual safety refresher and renewing in the member portal before March 1.

What you just ran

Each pair you write becomes one line of JSONL in the Alpaca format: an instruction (your question), an empty input, and an output (your answer). Unsloth reads that file and trains small QLoRA adapters on your pairs while the base model stays frozen. The PII scan is a first pass only: emails, US phone numbers, and the words "ssn" and "social security" get flagged, and then a person reads every pair anyway. This studio teaches the format of the training file, which is the first deliverable of the whole fine-tune.

The full tutorial walks you through eight steps: deciding what may train, writing 50 to 200 reviewed pairs, scanning for PII, holding out 10 pairs for evaluation, training on a free Colab GPU or a local one, testing the merged model against the base model, and picking a deployment route. Read the tutorial.

Download the starter kit

The demo above runs in your browser on fictional teaching pairs. The starter kit is the real thing: a spreadsheet template for your Q&As, a local converter that turns the spreadsheet into JSONL and re-runs the PII scan, the PII review checklist, a QLoRA starter notebook, ten held-out eval questions with a scoring rubric, and a setup guide. Free with your email.