Key Concepts
Terminology and concepts used throughout the Commissioned platform.
Fine-tuning
Fine-tuning takes a pre-trained large language model (LLM) and trains it further on your specific data. Unlike prompting (where you give the model instructions at runtime), fine-tuning permanently changes the model's weights so it inherently understands your domain, tone, and patterns.
Think of it like this:
- Prompting = giving someone a cheat sheet before an exam
- Fine-tuning = teaching someone the subject so they don't need a cheat sheet
Base model
The base model is the pre-trained LLM you start from. It already has broad knowledge of language, code, reasoning, and world knowledge. Fine-tuning adds your specific data on top of that foundation.
Commissioned supports base models from three providers:
- OpenAI — GPT-4.1 family
- Google — Gemini 2.5 family
- Open-source — Qwen 3 family (runs on Commissioned's GPUs)
Fine-tuned model
The fine-tuned model is the result of training. It combines the base model's general capabilities with the specific patterns from your data. You interact with it through the chat UI or API.
LoRA adapter
LoRA (Low-Rank Adaptation) is a technique that produces a small set of adapter weights instead of modifying the entire model. It's used for open-source models (Qwen 3 8B) and has two advantages:
- Fast training — minutes instead of hours
- Portable — download the adapter file and run it on your own hardware
The adapter is typically 50–200 MB and sits on top of the full base model at inference time.
Training job
A training job (or fine-tune job) is a single training run. When you upload data and click "Create a Custom Model," a job is created and progresses through these stages:
Queued → Validating files → In progress → Succeeded / FailedEach job produces one fine-tuned model.
Provider
The provider is the company or infrastructure running the base model:
| Provider | Models | Infrastructure |
|---|---|---|
| OpenAI | GPT-4.1, GPT-4.1 Mini | OpenAI's cloud |
| Gemini 2.5 Flash, Flash Lite, Pro | Google Cloud | |
| GPU | Qwen 3 8B | Commissioned's own GPUs |
Conversation
A conversation is a saved chat session between you and a fine-tuned model. Conversations are persisted and can be resumed later. Each model can have multiple independent conversations.
API key
An API key authenticates requests to the OpenAI-compatible API. Keys are created from your profile, shown once, and can be revoked at any time. They grant access to all your models.