01 · MODEL CHOICE
VerifiedBase, Instruct, and Reasoning
Choose a model by its starting behavior and the capability gap—not by a fixed row count.
- Base = raw pretraining
- Instruct = aligned
- Reasoning ⊂ Instruct possible
- Look at behavior, not the label
PREREQUISITE
This lesson is foundational; no prerequisites required.
The budget, window, and …A Base model is a pretrained starting point and is not required to follow instructions. An Instruct model has already been aligned for conversation and instruction following. Reasoning is not always a mutually exclusive class: a model can be both reasoning-capable and instruct-tuned.
For a narrow, structured domain adapter, Instruct is the default. Use RAG for frequently changing knowledge and constrained decoding when output shape must be guaranteed.
First thought
Use Instruct below 300 rows and Base above 1,000 rows.
Correction
There is no universal row threshold; task, quality, token volume, and starting capabilities must be evaluated together.
Decision rule
Narrow, structured assistant task → start from an Instruct checkpoint.
CONCEPT DEPTH
Read the same concept at three depths. Pick a level, switch instantly. Recommended starting point for university students: UNDERGRAD.
Base provides pretrained starting weights. Instruct can use SFT and preference optimization for instruction following. Reasoning behavior may be supported by supervised examples, distillation or RL; RL is not required in every case.