Models
Base vs Instruct Model
Also known as: base model, instruct model, chat model
A "base" model just continues text; an "instruct" (or chat) model has been further trained to follow instructions and behave like an assistant.
There are two versions of a language model. The base model is the raw result of training on lots of text, it just predicts what text comes next, so if you ask it a question it might continue with more questions instead of answering. The instruct (or chat) model has had extra training to actually follow your requests and act like a helpful assistant. The friendly chatbots you use are instruct models.
- Base model: the output of pretraining, a pure next-token predictor. Powerful, but it does not naturally follow instructions or chat; it just continues text.
- Instruct / chat model: a base model further trained to follow instructions and behave helpfully, via instruction tuning and often preference tuning (RLHF/DPO).
Why the distinction matters
- What you use: consumer assistants are instruct/chat models.
- What developers choose: base models are sometimes preferred as a starting point for custom fine-tuning; instruct models are ready to use.
- Behaviour: ask a base model a question and it may not answer it; the alignment steps are what make it usable.
The base vs instruct distinction marks the pre- vs post-alignment checkpoint. The base model results from self-supervised pretraining and models the data distribution (next-token likelihood) without instruction-following or safety shaping. Instruct/chat models add supervised instruction tuning and typically preference optimisation (RLHF/DPO), yielding zero-shot instruction-following, dialogue formatting (role structure), refusals and tone (Ouyang et al. 2022).
Implications: base models can be stronger raw substrates for bespoke fine-tuning (no prior alignment to overwrite) and are useful for research and completion tasks, while instruct models are deployment-ready and safer out of the box but carry the provider's behavioural priors. Prompting differs, base models need completion-style or few-shot framing, instruct models expect instructions in a chat template, and evaluation must specify which variant, since alignment substantially changes measured behaviour, calibration and refusal rates.