JOVANA
Explore Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

From Word-Guesser to Helpful Chatbot

A raw next-word predictor isn't yet an assistant. See how a base model becomes a chat model, and what "prompt" and "completion" really mean.

The raw model is a strange beast

Straight out of training, an LLM is a base model: a pure text-completion engine that has read the internet but has no idea it's supposed to be helpful. Ask a base model "What is the capital of France?" and it might answer "Paris" — or it might continue with three more quiz questions, because in its training data, questions are often followed by more questions. It completes text; it doesn't yet assist.

A base model is powerful but unruly. It has the raw knowledge, yet it doesn't reliably follow instructions, stay on task, or refuse harmful requests. Turning that raw capability into something you'd want to chat with takes a second stage of shaping.

A base model is just this loop: predict the next token, append it, repeat — pure text completion.

Diagram of an autoregressive generation loop feeding each predicted token back in as input.

Teaching it to be an assistant

The version you actually talk to is a chat model, also called a conversational assistant. It is built by taking the base model and further training it on examples of helpful dialogue — a step called instruction tuning — and then refining it from human feedback so it prefers answers people find useful and well-behaved. That feedback step often uses a method called reinforcement learning from human feedback (RLHF).

The base model's skills don't get erased — they get aimed. After this shaping (the broader practice is called fine-tuning), the same next-word machine now reliably answers your question, follows your formatting requests, asks for clarification, and declines things it shouldn't do. That is why the assistant you use feels so much more cooperative than its raw ancestor.

Turning a base model into a helpful assistant: human preferences train a reward model that tunes the policy.

RLHF pipeline: human preference data feeds a reward model that fine-tunes the language model's policy.

Prompt in, completion out

Two words you'll see everywhere: the prompt is everything you send the model — your question plus any earlier conversation and instructions — and the completion is the text it writes back. Internally there is no magic line between them: the model simply continues the prompt, and the part it adds is the completion. Your whole conversation is one growing piece of text that the model keeps completing, turn after turn.

PROMPT (what you and the system provide):
  System: You are a helpful, concise assistant.
  User:   What's the capital of France?

COMPLETION (what the model writes):
  Assistant: The capital of France is Paris.
A single turn: the prompt sets the scene and asks; the completion is the model continuing that text as the assistant.

Why chat models feel like magic (and what's underneath)

It's worth saying plainly: even the polished assistant is still running next-token prediction underneath every reply. The fine-tuning didn't replace that loop; it taught the loop to land on helpful, safe, well-formatted continuations. When the chatbot "decides" to ask you a clarifying question, that decision is just the model predicting that a clarifying question is the most fitting next text, given how it was trained to behave.

p(x_t \mid x_{<t}) = \mathrm{softmax}(W h_{t-1})

Even the polished chatbot is still choosing each token from this next-token distribution.

Holding both truths at once — a humble next-word predictor and a genuinely useful collaborator — is the mature way to understand these systems. It keeps you from over-trusting them while still letting you get real work done.