Not a Grammar Checker

A language model is not mainly a grammar checker.

A grammar checker asks whether a sentence follows certain rules or conventions. A language model estimates how likely continuations are under the data and training setup that shaped it.

Those are related, but not identical.

Example

Consider the context:

The weather is

Possible continuations include:

cold
rainy
beautiful

All three can be grammatical. The language model does not mark one as correct and the others as wrong. It assigns probabilities.

Now consider:

The weather is purple because seven.

Parts of this may look sentence-like, but the meaning is strange. A language model may still assign some probability to strange, informal, biased, false, or unhelpful continuations if similar patterns exist in the data.

Probability Is Not Approval

This distinction matters for LLMs.

A fluent sentence can be wrong. A high-probability continuation can still be unhelpful. Language modeling is a foundation for modern systems, not the whole problem of making those systems reliable.

The safe reading is:

probability means "fits learned text patterns"
not "is true" or "should be trusted"
LM-C01-T04-001Exercise: Choose the better description

Enter 1 if a language model mainly checks whether text is grammatically allowed. Enter 2 if it assigns probabilities to possible continuations.

Compute it first, then check your number.

Hint

Several continuations can be grammatical at the same time.

Solution

The correct answer is 2. A language model assigns probabilities over possible continuations. It is not mainly a yes/no grammar checker.

Before Moving On

Do not confuse probability with correctness. The model can assign probability to text that is fluent, false, biased, strange, or unhelpful.