PopStory!!

Corrections

If you read one of these languages properly and you find a mistake, we would like to know. It takes a minute, you do not need an account, and you will not get a newsletter.

Why we are asking

Language models are good at small languages in the way a confident learner is good at them: fluent enough to sound right, and wrong in ways a first language speaker spots instantly. In Welsh it shows up as mutations, or northern and southern forms mixed inside one paragraph. In Irish it is one dialect's grammar with another's vocabulary. In Latin it is English word order wearing Latin endings. The output reads well, which is precisely what makes it dangerous for someone learning from it.

This is a data problem, not a mystery. These models have seen an enormous amount of English and very little of everything else, and nobody has assembled the kind of material that shows exactly where they go wrong. That gap is why smaller languages are handled badly, and it does not close on its own.

What happens to a correction

Every story here is stored with the exact instruction that produced it and the model that wrote it. When someone tells us a sentence is wrong and what it should say, that becomes a complete record: this model, given this instruction, produced this, and someone who reads the language says it should have been that.

We publish those records openly and free of restrictions, so anyone building or testing a model can use them. For the languages that need it most, that is worth considerably more than the stories are.

Corrections are credited if you want them to be and anonymous if you do not. You do not have to approve of any of this to send one. If you think generated writing is a bad idea and you also think our Welsh is wrong, both can be true, and the second one is useful to us.

How to send one

Copy the sentence, say what is wrong with it and what it should be, and include the address of the page. That is all we need. A short note is better than nothing: if you can only tell us which sentence is wrong and not why, send that.

Send a correction