Poro 2 8B Base is an 8B parameter decoder-only transformer created through continued pretraining of Llama 3.1 8B to add Finnish language capabilities. It was trained on 165B tokens using a carefully balanced mix of Finnish, English, code, and math data. Poro 2 is a fully open source model and is made available under the Llama 3.1 Community License.

Poro 2 was created in a collaboration between AMD Silo AI, the TurkuNLP group of the University of Turku, and High Performance Language Technologies (HPLT). Training was conducted on the LUMI supercomputer, using compute resources generously provided by CSC - IT Center for Science, Finland.

This model demonstrates how continued pretraining can efficiently add new language capabilities to existing models while maintaining performance in the original domains. Through the combination of English and Finnish training data, we achieve a model that substantially outperforms the base Llama 3.1 8B model in Finnish while maintaining solid English proficiency.

For more details on our training and data curation process, check out our Continued Pretraining Playbook.

Poro 2 Model Family

The Poro 2 model family includes both 8B and 70B models, and there are three different versions released of the Poro 2 models: a base model, a post-training SFT-only checkpoint, and the final instruct model which is the SFT model plus a round of DPO.

Model	Based on	Base Model	SFT	Instruct
Poro 2 8B	Llama 3.1 8B	Poro 2 8B Base	Poro 2 8B SFT	Poro 2 8B Instruct
Poro 2 70B	Llama 3.1 70B	Poro 2 70B Base	Poro 2 70B SFT	Poro 2 70B Instruct

What does Poro mean? Poro is the Finnish word for Reindeer! 🦌 These animals are native to Finland and hold a significant and historical role in Finnish culture.

Model Overview

NOTE: This is a base model which needs further fine tuning for most use cases.

Poro 2 8B is based on the Llama 3.1 8B architecture and uses continued pretraining to add Finnish language capabilities.

Hyperparameter	Value
n_parameters	8.03B
n_layers	32
n_heads	32
n_kv_heads	8
d_model	4096
vocab_size	128256
max_sequence_length	8192
base_model	Llama-3.1-8B

Training

Poro 2 8B was created through continued pretraining on the LUMI supercomputer, using AMD MI250X GPUs. Training used a 3D parallelism strategy with TP=2, PP=1.

Training was conducted using a custom version of the Megatron-LM framework. Our code is available at https://github.com/LumiOpen/Megatron-LM-lumi.

Training Hyperparameters

Hyperparameter	Value	Comment
Precision	bfloat16
Optimizer	AdamW
Learning rate	3e-4
LR scheduler	cosine	Warmup ratio 0.05, min LR 1e-8
Weight decay	1e-1
Global batch size	512
Micro batch size	1
Max sequence length	8192
Total tokens	165B	1 epoch

Dataset

Poro 2 8B was trained on a balanced 165B token dataset designed to maintain English, code, and math capabilities while adding Finnish proficiency.

Dataset	Source	Percentage	Tokens
Finnish	FineWeb2	30%	50B
English	FineWeb-Edu	30%	50B
Code	StarCoder	30%	50B
Math	FineMath	10%	16B
Total		100%	165B

Evaluation Results

Poro 2 8B shows substantial improvements in Finnish capabilities over Llama 3.1 8B, while maintaining English performance:

Finnish Performance

	Poro 2 8B	Llama 3.1 8B
ARC Challenge	48.90	38.82
HellaSwag	50.49	30.97
MMLU	56.25	49.64
TruthfulQA	49.78	45.54

English Performance

	Poro 2 8B	Llama 3.1 8B
ARC Challenge	60.75	57.94
HellaSwag	80.55	80.05
MMLU	63.48	65.08
TruthfulQA	48.06	54.02

Translation Performance

	Poro 2 8B	Llama 3.1 8B
EN→FI BLEU	36.48	23.92
FI→EN BLEU	40.71	37.42

Overall: ~10 percentage point average improvement in Finnish benchmarks with only ~1 percentage point decrease in English performance.

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "LumiOpen/Poro-2-8B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Example usage
prompt = "Kerro minulle Suomesta."  # "Tell me about Finland" in Finnish
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Ethical Considerations and Limitations

Poro 2 8B is an advanced language model optimized for English and Finnish, with additional capabilities in code and mathematics. As with most AI-driven systems, Poro 2 is a product of the vast data it has been trained on, which may reflect the imperfections, biases, and idiosyncrasies of the wider web. The model may, at times, produce outputs that can be considered inaccurate, prejudiced, or controversial.

Key limitations:

Limited proficiency in languages other than English and Finnish
Potential for generating biased or inappropriate content
May produce factually incorrect information

Users and developers engaging with Poro 2 should exercise discretion and consider additional evaluation and customization to ensure the model's responses align with their specific needs.

License

Built with Llama

Poro 2 8B is released under the Llama 3.1 Community License. Please review the license terms before use.

Citation

bibtex

@misc{poro2_2025,
    title={Poro 2: Continued Pretraining for Language Acquisition},
    author={Elaine Zosa and Jouni Luoma and Kai Hakala and Antti Virtanen and Mika Koistinen and Risto Luukkonen and Akseli Reunamo and Sampo Pyysalo and Jonathan Burdge},
    year={2025},
    howpublished={LumiOpen}
}

Acknowledgments

We thank CSC - IT Center for Science, Finland for providing access to the LUMI supercomputer. This work was supported by the High Performance Language Technologies (HPLT) project and conducted in collaboration with TurkuNLP from the University of Turku. This project has received funding from the European Union's Horizon Europe research and innovation programme under grant agreement No 101070350.

LumiOpen/Llama-Poro-2-8B-base

Poro 2 8B Base Model Card