AMÁLIA and the future of European Portuguese LLMs (opens in new tab)

Given that their publication says the dataset is freely available on Huggingface that's at least something ig

Why?

You are right about most tokenizers being heavily biased towards English, but the situation is not so bad for Portuguese. Here are some results on the Goldfish corpus [1] with a few different tokenizers. This measures #characters in corpus / #subwords in tokenized corpus.

```

Llama3

english, 0.216

portuguese, 0.285

italian, 0.287

greek, 0.592

```

```

Gemma4

english, 0.219

portuguese, 0.246

italian, 0.249

greek, 0.537

```

```

Kimi2.6

english, 0.214

portuguese, 0.310

italian, 0.308

greek, 0.716

```

Portuguese is worse than English certainly, but it is on par with Italian (which I think has more overlap with English) and much better than Greek (since it doesn't use the Latin script and is definitely not prioritized in the tokenizer construction).

On your second point, tokenizer transfer allows for extending/modifying a tokenizer without retraining the model from scratch. The simplest version of this is tokenizer extension + continual pretraining, where you just add a bunch more tokens to the vocab for the language/domain that you want to improve and train a little more. It's been done for Japanese [2] and Indic languages, but afaik not Portuguese.

So I think that continual pretraining for a large base model would have probably been fine for this case with huge cost savings. But it is good to have the ability to train your own base models, so I don't think this is such a bad idea.

-----------------------

[1]: https://huggingface.co/datasets/goldfish-models/fish-food

[2]: https://arxiv.org/abs/2404.17790

I just died.

The authors are pretty clearly trying to draw only from European Portuguese sources - I feel like there's a fairly widespread attitude here that the language is being overwhelmed by the sheer number of Brazilian speakers (which there is obviously at least some truth to).

I don't necessarily personally feel like preserving European Portuguese in amber is a worthwhile goal (anymore than it is productive for Brits to be prickly about the meteoric rise of US English)

Right, but most of those speak brazilian portuguese. There's so much less european portuguese text that it becomes impossible for a model to not speak brazilian portuguese if not trained in a way that ignores brazilian sources

The whole point of this project is to have an LLM that speaks European Portuguese, not Brazilian Portuguese.

Mutually intelligible, yes, but far from perfectly so. I speak both, as a native anglophone, and the difference is not so much “US vs British English” so much as “Guyanese English vs British English”. Like, fundamental points of grammar differ, the spoken rhythm and syllabic stress differs (poetry does not translate well between them), never mind just vocabulary. Continental Portuguese people tend to find it easier to understand brasileiros than vice versa, largely due to mostly one-way cultural exports, but to try to roll both into a single model would create a creole at best.

> European Portuguese is the 13th most populous language in Europe

that's not impressive

To my knowledge, all major LLMs are multilingual. This article could really have used an evaluation of existing models' European Portuguese capabilities.

yeah, they seem all confined to being an American-consultant-Chinese-authoritarian split personality with broad second language capabilities. I suppose they become too incoherent otherwise.

E.g. gemma3:4b can fake simple conversations in several european languages, including portuguese, swedish and finnish.

It's just a database. If you push text in one language into it, it'll likely crap out stuff in that same language, unless the system prompt that also goes in with your query causes it not to.