alt Hacker NewsI'm not arguing with the rest of your points, but...
> Just fine tune the open source Kimi 2.6 with an open source Portuguese dataset
I think that tokenizers of all popular models are heavily biased towards English or English and Mandarin.
And I don't think that it is possibple to replace the tokenizer without full retraining.
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