Here's a "continued pre-trained" model using Finnish Wikipedia dataset. I still don't understand why no one in Finland has figured out that they could just do continued pre-training on existing models that are already supported by every frontend.. I've seen Japanese models perform pretty well with that kind of continued pre-training, yet Finnish models are still done from scratch which means they suck ass. If you compare them to Llama 3 or Gemma 2 they just suck so much. They can't even match Mistral 7B a model from last year. Just stop wasting money on training models from scratch, use these better models as base and train it on all your closed-source data I don't have access to. Thank you.

LoRA: mpasila/Llama-3.2-Finnish-Wikipedia-LoRA-1B

Trained with regular LoRA (not quantized/QLoRA) and LoRA rank was 128 and Alpha set to 32. Trained for 1 epoch using RTX 4090 for about 12,5 hours.

So it does have some issues but I could try training it on Gemma 2 2B and see if that's a better model for this (Gemma 2 already is better at Finnish than Llama 3) and maybe add more datasets containing Finnish.

Evaluation

Source

FIN-bench scores:

Uploaded Llama-3.2-Finnish-Wikipedia-1B model

• Developed by: mpasila
• License: Llama 3.2 Community License Agreement
• Finetuned from model : unsloth/Llama-3.2-1B

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.