Base model versus assistant
The model that continues text and the product that answers your message are related, but they are not the same thing.
In this lesson
Assistant behavior comes from learned adaptation and a surrounding system, not from an interface label alone.
The checkpoint and the experience
A base model is typically the result of broad pretraining before assistant-oriented adaptation. It can predict many styles of text, including dialogue, yet continuing a document and responding helpfully to a user are different behavioral targets. An assistant-tuned model has undergone additional training intended to make useful conversational behavior more likely.
The 2023 Llama 2 report explicitly distinguishes pretrained models from fine-tuned chat variants. This is a dated example of the distinction, not a statement that all model families publish both versions or use identical adaptation. Always inspect the actual model documentation when identifying a checkpoint.
Why the same prompt can lead elsewhere
Imagine the input “Explain tides.” A text-completion model could continue a worksheet, invent the next instruction, or supply an explanation. Those are all conceivable continuations of a document. Assistant training and a conversational format make answering the request a more consistent target. They do not turn probabilistic generation into a deterministic fact lookup.
A base model may also follow instructions through suitable prompting. The boundary is not “unable versus able.” It is a difference in training history and the distribution of behaviors expected under particular inputs. Capability, instruction following, and product safeguards should be evaluated separately.
The application adds another layer
A chat application may assemble system instructions, user messages, tool descriptions, retrieved documents, and selected conversation history before invoking a model. It may execute a requested tool, feed the result back, or apply additional policies. The visible answer can be the product of several operations around one or more model calls.
None of these features follows automatically from the term LLM. A local text model with no network tool cannot secretly obtain live web results merely because the prompt asks for them. A system that can retrieve documents still needs an actual retrieval step and a way to provide those documents as context.
A useful way to investigate behavior
When an assistant gives a surprising answer, ask which layer could explain it. Was the model trained for this kind of response? Was an instruction present? Did the application truncate an earlier message? Did a tool fail or return stale information? These questions are more precise than assuming every behavior lives in the weights.
This separation also helps interpret comparisons. Evaluating a bare checkpoint against a tool-enabled product mixes model quality with system capabilities. A fair test states which model, prompt construction, tools, sampling settings, and evaluation criteria were used. The combination is often what users care about, but its ingredients should remain visible.
One request, three layers
“What is the current train delay?” requires language understanding, access to a current timetable source, and application logic that requests and presents the result. A well-tuned conversational model supplies only part of that chain.
A chatbot is simply a base model with a text box.
A deployed assistant may add post-training, message formatting, retrieval, tools, and policy around the underlying model.
Take it with you
- Distinguish pretraining from assistant adaptation.
- Tool access belongs to the surrounding system.
- Compare complete setups, not only model names.
Check your understanding.
Choose an answer, then see the explanation. You can try again.
Practice only. Your answer is not stored or sent anywhere.
Prompt structure: what the model receives
Continue when you’re ready. Marking this lesson as read is optional.
Sources & further reading
Primary references for this lesson. The explanation and examples above are original.
- Llama 2: Open Foundation and Fine-Tuned Chat Models ↗Hugo Touvron et al. · 2023-07-18 · checked 13 Nov 2026