Llama Poro 2 8B Instruct

🚀 Poro 2 8B Instruct Model Card

Poro 2 8B Instruct is an instruction-following chatbot model. It's developed by applying supervised fine-tuning (SFT) and Direct Preference Optimization (DPO) to the Poro 2 8B Base model. This model is tailored for conversational AI applications and instruction following in both Finnish and English. It was trained on a well - curated mix of English and Finnish instruction data, followed by preference tuning to enhance response quality.

Poro 2 was created through a collaboration among AMD Silo AI, the TurkuNLP group of the University of Turku, and High Performance Language Technologies (HPLT). The training was carried out on the LUMI supercomputer, using compute resources provided by CSC - IT Center for Science, Finland.

This model showcases how continued pretraining followed by instruction tuning can effectively add new language capabilities to existing models while maintaining strong conversational skills in both the original and target languages.

For more details about our training and data generation pipeline, check out our Continued Pretraining Playbook.

✨ Features

Poro 2 Model Family

The Poro 2 model family includes 8B and 70B models. There are three different versions 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.

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

Model Overview

Poro 2 8B Instruct is based on the Llama 3.1 8B architecture and has been fine - tuned for instruction following and conversational AI applications. The model supports both English and Finnish conversations.

🔧 Technical Details

Training Process

Continued Pretraining

The base Poro 2 8B model was created through continued pretraining on 165B tokens of Finnish, English, code, and math data.

Supervised Fine - Tuning (SFT)

The SFT phase used 1.4M instruction - following examples in English and Finnish, including:

• English and Finnish Tulu 3 prompts with Llama - 3.3 - 70B - Instruct responses
• Multi - turn conversations generated using the Magpie method
• Top - rated conversations from OASST2 and Avoin Avustaja datasets
• Translation samples from EuroParl

We also release the Poro 2 instruction collection.

Direct Preference Optimization (DPO)

The final model underwent preference tuning using the HelpSteer3 dataset to improve response quality and alignment.

Post - Training Hyperparameters

SFT

DPO

Evaluation Results

Poro 2 8B Instruct shows substantial improvements in Finnish instruction - following capabilities compared to Llama 3.1 8B Instruct, while maintaining strong English performance.

Finnish Instruction Following

English Instruction Following

Pairwise Comparisons (MTBench)

• Finnish: 85% win rate vs Llama 3.1 8B Instruct
• Finnish: 51% win rate vs Llama 3.3 70B Instruct
• English: 49% win rate vs Llama 3.1 8B Instruct

Overall: ~24% average improvement in Finnish instruction - following benchmarks while maintaining English performance.

💻 Usage Examples

Basic Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

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

# Finnish conversation example
messages = [
    {"role": "user", "content": "Kerro minulle Suomen historiasta."}
]

inputs = tokenizer.apply_chat_template(
    messages, 
    add_generation_prompt=True,
    return_tensors="pt"
)

outputs = model.generate(
    inputs,
    max_new_tokens=500,
    temperature=0.7,
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

📚 Documentation

Intended Use

Poro 2 8B Instruct is designed for:

• Conversational AI applications in Finnish and English
• Question answering and information retrieval
• Content generation and creative writing
• Educational applications
• Customer service and support applications
• Translation between Finnish and English

Ethical Considerations and Limitations

Poro 2 8B Instruct is an advanced conversational AI model optimized for English and Finnish instruction following. As with most AI - driven systems, this model may reflect imperfections, biases, and idiosyncrasies present in its training data.

Key limitations:

• Limited proficiency in languages other than English and Finnish
• May occasionally generate biased, inappropriate, or factually incorrect content
• Performance may vary significantly for specialized or technical domains
• Context window limited to 8,192 tokens
• May struggle with very recent events (knowledge cutoff limitations)

⚠️ Important Note

Users should verify important factual claims independently. The model should not be used for medical, legal, or financial advice without human oversight. Responses should be reviewed for appropriateness in sensitive contexts.

📄 License

Built with Llama

Poro 2 8B Instruct is released under the Llama 3.3 Community License. Please review the license terms before use.

Citation

@misc{poro2_2025,
    title={Poro 2: Continued Pretraining for Language Acquisition},
    author={Elaine Zosa and Jouni Louma 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.