Commissioned
Getting Started

Your First Model

A detailed walkthrough of uploading data and creating your first fine-tuned model.

Before you start

Have your data ready. If you're just experimenting, any of these will work:

  • A few blog posts or essays you've written (.txt or .md)
  • Chat or email transcripts (.jsonl or .txt)
  • Documentation or notes (.pdf or .md)
  • Code files (.txt)

Don't worry about formatting. Commissioned handles cleaning, structuring, and deduplicating your data automatically. Raw, messy files are fine.

Step-by-step

Upload files

Drag and drop files into the upload area, or click to browse. You can upload multiple files at once.

Supported formats: JSONL, JSON, PDF, TXT, Markdown

Size limit: 5 GB per file

The upload card shows a preview of your selected files with names and sizes. You can remove individual files before submitting.

Write your use-case description

This is the most important input besides your data. The description tells Commissioned how to structure your data for training.

Good descriptions are specific:

❌ Vague✅ Specific
"Make a chatbot""Create a customer support agent that answers questions about our SaaS product using the attached help docs and past ticket transcripts"
"Writing helper""I want a writing assistant that matches the academic tone and citation style in these research papers"
"Code assistant""Fine-tune a coding assistant that follows our team's TypeScript conventions and knows our internal API from these docs"

Select a base model

Choose which LLM to fine-tune. See Models for a detailed comparison.

Quick recommendations:

  • First time? → GPT-4.1 Mini (fast, free, solid quality)
  • Need the best quality? → GPT-4.1 or Gemini 2.5 Pro
  • Want to self-host? → Qwen 3 8B (produces a downloadable LoRA adapter)
  • Have very long documents? → Gemini 2.5 Flash

Submit and wait

Click Create a Custom Model. Your model card appears in the left panel with real-time status updates:

StatusWhat's happening
QueuedJob is waiting to start
Validating filesData is being parsed and cleaned
In progressModel is actively training
SucceededDone — ready to use
FailedSomething went wrong

Training times:

  • GPU / Qwen: ~5 minutes
  • OpenAI / Gemini: 30–45 minutes

You'll receive an email notification when training completes.

After training completes

Once your model shows Succeeded, you can:

  1. Chat with it — click Open in chat to start a conversation right in the browser
  2. Call it via API — generate an API key and use the OpenAI-compatible endpoint
  3. Rename it — click the rename action on the model card for better organization
  4. Download the adapter — for Qwen models, download the LoRA weights to self-host

Troubleshooting

Model shows "Failed"

Common causes:

  • Not enough usable text — your files need enough content for training (a few hundred lines minimum)
  • Corrupted file — try re-exporting your PDF or converting to plain text
  • Provider issue — transient errors happen; create a new job with the same data

Training is taking longer than expected

Cloud models (OpenAI, Gemini) occasionally queue during high demand. If it's been over 2 hours, try creating a new job. GPU models (Qwen) should always complete within 10 minutes.

Model responses don't match my data

  • Add more data — fine-tuning works better with more examples
  • Be more specific in your use-case description
  • Try a different base model — some models pick up certain patterns better than others

On this page