Instruct model

Programs like ChatGPT initially learn only a single task: predicting, from vast amounts of text, which word comes next. Such a raw model is astonishingly articulate, but it obeys no one. If you write to it “Explain photosynthesis to me,” it might respond with further exam questions, because in its training texts such sentences were often followed by more tasks. An instruct model is the revised version of such a raw model. In a second round of training, it has seen thousands of examples in which an instruction is followed by a fitting, helpful answer. After that, it understands a request as a request and delivers what is wanted.

From text continuer to usable assistant

The difference between a raw model and its instruct variant is enormous in everyday use. The raw model possesses the same knowledge, but you have to coax it out with tricks. You have to start the text in such a way that the desired answer is the most natural continuation. Experts can manage this, ordinary users cannot. It was only instruction training that made language models usable for the general public.

This is exactly where a common misconception lies. Many believe the chatbots became so good because they learned more and more. But a large part of the progress did not come from more knowledge, but from better obedience. The actual capability was already present in the model beforehand. It just wasn’t accessible.

For companies, the instruct variant is therefore usually the only one that comes into question at all. Anyone building AI into customer service needs reliable answers to clear requests. A raw model would react unpredictably there. Providers therefore often release their models in two versions, recognizable by name additions like “Base” and “Instruct.”

What happens during instruction training

The first stage is called Instruction Tuning. Humans write or collect pairs of instruction and sample answer for this. One pair might read: “Summarize this text in three sentences,” followed by a good summary. The model continues training on these examples and adopts the pattern. It doesn’t learn new facts, but a new way of interacting.

This is usually followed by a second stage involving feedback from humans. Test subjects are shown several answers from the model and rate which one is better. The model is fine-tuned based on these ratings. This procedure is often referred to by the abbreviation RLHF, which stands for Reinforcement Learning from Human Feedback. It ensures that answers not only fit formally, but are also polite, safe, and to the point.

This second training step is tiny compared to the first. The base training consumes months and enormous computing costs. Instruction training often gets by with a few tens of thousands of examples. You can picture it like a person who has read for years and then, in a few weeks, learns to conduct themselves in a consulting conversation. The knowledge is old, the behavior is new.

How to recognize instruct models on offer

They are most clearly visible in the model names. Designations like “Llama 3 Instruct” or “Mistral Instruct” point directly to it. On platforms where freely available models can be downloaded, you almost always find both versions side by side. Anyone experimenting on their own should reach for the instruct version, unless they plan their own training.

In the news, the term usually surfaces when reporting on the costs or safety of AI. Safety rules, such as refusing dangerous requests, are mostly built in during this training phase. Criticism of an AI that answers too cautiously or too verbosely is therefore almost always directed at the instruction training. The raw model underneath may have remained completely unchanged.

A related term is the chat model. This refers to an instruct variant that is additionally designed for multiple consecutive rounds of conversation. It thus remembers the course of the conversation. In practice, both words are often used synonymously, even though that is not entirely accurate technically.