Chat Model

What is a Chat Model?

A Chat Model is created when a Foundation Model is fine-tuned with example conversations to follow instructions, respond in a particular style, and answer questions directly instead of merely continuing text. Nearly all publicly available chat assistants today are based on Chat Models.

How it works

After pretraining, a Chat Model undergoes additional fine-tuning using example conversations that demonstrate how to respond to instructions and questions. Special control tokens or chat templates identify the roles of the system, user, and model within the context.

Why a Chat Model enables conversation

Without this additional fine-tuning, a model would often respond to a question by continuing the text plausibly instead of providing a direct answer. Only Chat Model fine-tuning makes a language model significantly better suited for conversational applications.

A Chat Model does not automatically possess more knowledge than the underlying Foundation Model. Fine-tuning primarily changes response behavior and style rather than the model’s knowledge. A Chat Model may even become less flexible through fine-tuning if the training data followed highly uniform structures.

Why the Chat Model has practical implications

• Everyday use: Nearly all publicly used language model applications are based on Chat Models rather than pure Foundation Models.
• System Prompt influence: Chat templates and role markers influence how strongly a System Prompt affects the model’s behavior.
• No increase in knowledge: Switching from a Foundation Model to a Chat Model changes response behavior, not automatically factual accuracy.

Typical example

Given the same prompt, “Explain Docker”, a Chat Model provides a direct explanation, whereas a pure Foundation Model would more likely continue the prompt as the beginning of a text.

Further reading

Fundamentals

→ How a Language Model Works

Related terms

→ Foundation Model
→ System Prompt