Foundation models concept
Base versus instruct models
ReferenceAn instruct model is a language model adapted after pretraining to respond to instructions, often in a conversational format. A base model is the earlier pretrained model, shaped mainly to continue text; it may not treat a question as a request for a helpful answer. “Base versus instruct” is mainly a difference in training and expected behavior, not a guarantee that one knows more, is more reliable, or fits every task. The labels overlap with “chat model,” though they describe related rather than perfectly identical ideas.
Reviewed 2026-10-09
The two kinds of model
A base model is a language model after its broad initial training, often called pretraining. It has learned patterns in text and can continue a passage, but it has not necessarily been adapted to handle chat or follow requests. An instruct model starts from a pretrained model and receives further training on examples of instructions and suitable responses. Some approaches also use human rankings of model responses as a training signal. The shared starting point matters: “instruct” describes additional training and intended behavior, not a completely separate kind of model.
What changes in practice
The practical difference is the model’s expected response to an input. A base model is generally used to continue text, while an instruct model is trained to respond to directions such as “summarize this” or “explain the term.” Chat models commonly accept conversations as messages with roles and content, and many are also called instruct or instruction-tuned models. The training format matters too: a chat template arranges messages in the form expected by the model. These labels describe tendencies, not a promise that an instruct model will follow every request or give a correct answer.
When each may fit
An instruct model is a practical starting point when you want a model to answer questions or follow everyday written directions, because instruction-following behavior is what its additional training targets. A base model may fit when your task is text continuation, or when you plan to adapt the model yourself for a particular use. For example, supervised instruction tuning has been used to adapt a pretrained model with instruction-and-response demonstrations. The right choice still depends on the task and the model’s documented behavior; the label alone does not tell you how well it will perform.
Common confusions
“Base model vs chat model” is close to “base vs instruct,” but the labels describe different angles. Base and instruct refer mainly to training stage and behavior; chat describes a conversational way to interact with a model. In practice, the terms overlap because many chat models are instruction-tuned. A chat interface or a formatted conversation does not by itself show how a model was trained. Also, instruction tuning is not the same as a guarantee of accuracy: research on instruction-following models reports that they can still fail to follow directions or make up facts.
Questions people ask
Base model vs chat model: are they different?
They describe related but not identical things. “Base” usually means a pretrained model that does not support chat; “chat” describes a conversational model and its message-based interaction. Many chat models are also called instruct models.
Is an instruct model always trained with human feedback?
No single recipe defines the label. Instruction-following models have been trained on demonstrations of desired responses, and some approaches further tune a model using human rankings of its outputs.
Will an instruct model always answer more accurately?
No. Instruction tuning targets how a model responds to directions; it does not certify that its answers are true. Research on instruction-following models describes failures such as making up facts and not following instructions.
Can a base model be adapted for a specific task?
Yes. A pretrained model can be further tuned using examples for a task, including demonstrations that pair instructions with responses.
Learn more
- Chat basics: chat models and message formats
Explains how chat models use messages and why “chat” and “instruct” are often overlapping labels.
- Training language models to follow instructions with human feedback
Gives a research example of instruction demonstrations followed by tuning from human rankings.
- How supervised instruction tuning works
Shows an example of adapting a pretrained model with instruction-following demonstrations.
What changed
Sources
Part of Foundation models: What kind of model is it?How adaptation changes a model