Poro 2 70B SFT Model Card
Note for most users: This is an intermediate checkpoint from our post-training pipeline. Most users should use
Poro 2 70B Instruct instead, which includes an additional round of Direct Preference Optimization (DPO) for improved response quality and alignment. This SFT-only model is primarily intended for researchers interested in studying the effects of different post-training techniques.
Poro 2 70B SFT is a supervised fine-tuned model created from the Poro 2 70B Base model. This model has been trained for instruction following and conversational AI applications in both Finnish and English, but has not undergone preference tuning. It represents the intermediate step before Direct Preference Optimization (DPO) in our post-training pipeline.
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.
For more details on our training and data generation pipeline, 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.
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
Poro 2 70B SFT is based on the Llama 3.1 70B architecture and has been supervised fine-tuned for instruction following. The model supports both English and Finnish conversations but has not undergone preference tuning for response quality optimization.
| Hyperparameter | Value |
|---|
| n_parameters | 70.55B |
| n_layers | 80 |
| n_heads | 64 |
| n_kv_heads | 8 |
| d_model | 8192 |
| vocab_size | 128256 |
| max_sequence_length | 8192 |
| base_model | Llama-3.1-70B |
Training Process
Continued Pretraining
The base Poro 2 70B model was created through continued pretraining on 165B tokens of Finnish, English, code, and math data.
Supervised Fine-Tuning (SFT)
This model represents the SFT phase of post-training, using 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
SFT Hyperparameters
| Hyperparameter | Value |
|---|
| Epochs | 2 |
| Global batch size | 128 |
| Learning rate | 5e-6 |
| LR scheduler | linear |
| Warmup ratio | 0.03 |
| Max sequence length | 4,096 |
Evaluation Results
Poro 2 70B SFT shows substantial improvements in Finnish instruction-following capabilities compared to Llama 3.1 70B Instruct and is on par with Llama 3.3 70B Instruct, while maintaining excellent English performance. Note that the final Instruct model (with DPO) performs better.
Finnish Instruction Following
Finnish Instruction Following
| Poro 2 70B SFT | Llama 3.1 70B Instruct | Llama 3.3 70B Instruct | Poro 2 70B Instruct |
|---|
| IFEval Finnish | 70.05 | 63.95 | 71.71 | 70.79 |
| MTBench Finnish | 7.2 | 7.06 | 7.4 | 7.77 |
| AlpacaEval 2 Finnish | 30.74 | 21.06 | 25.73 | 41.96 |
English Instruction Following
| Poro 2 70B SFT | Llama 3.1 70B Instruct | Llama 3.3 70B Instruct | Poro 2 70B Instruct |
|---|
| IFEval | 89.46 | 86.69 | 90.38 | 85.95 |
| MTBench | 8.03 | 8.33 | 8.35 | 8.41 |
| AlpacaEval 2 | 43.18 | 43.87 | 45.12 | 49.77 |
MTBench scores per-category
| Finnish | English |
|---|
| Coding | 6.50 | 7.15 |
| Extraction | 7.95 | 8.95 |
| Humanities | 8.50 | 8.80 |
| Math | 6.85 | 8.35 |
| Reasoning | 5.20 | 6.55 |
| Roleplay | 7.60 | 8.30 |
| STEM | 7.25 | 7.80 |
| Writing | 7.80 | 8.35 |
MTBench scores per-turn
| Finnish | English |
|---|
| first turn | 7.83 | 8.55 |
| second turn | 6.57 | 7.51 |
Overall: Notable improvement over Llama 3.1 70B Instruct and competitive with Llama 3.3 70B Instruct in Finnish, while maintaining strong English performance. The additional DPO step in the Instruct model provides further improvements.
Usage
1from transformers import AutoTokenizer, AutoModelForCausalLM
2import torch
3
4model_name = "LumiOpen/Poro-2-70B-SFT"
5tokenizer = AutoTokenizer.from_pretrained(model_name)
6model = AutoModelForCausalLM.from_pretrained(
7 model_name,
8 torch_dtype=torch.bfloat16,
9 device_map="auto"
10)
11
12# Finnish conversation example
13messages = [
14 {"role": "user", "content": "Kerro minulle Suomen historiasta."}
15]
16
17inputs = tokenizer.apply_chat_template(
18 messages,
19 add_generation_prompt=True,
20 return_tensors="pt"
21)
22
23outputs = model.generate(
24 inputs,
25 max_new_tokens=500,
26 temperature=0.7,
27 do_sample=True,
28 pad_token_id=tokenizer.eos_token_id
29)
30
31response = tokenizer.decode(outputs[0], skip_special_tokens=True)
32print(response)
Research Applications
This SFT-only model is particularly useful for researchers studying:
- The effects of supervised fine-tuning vs. preference tuning
- Comparative analysis of different post-training techniques
- Ablation studies on instruction-following capabilities
- Cross-lingual transfer in instruction-following tasks
- The impact of DPO on model behavior and alignment
Intended Use
Poro 2 70B SFT is primarily intended for:
- Research purposes: Studying post-training techniques and their effects
- Comparative analysis: Understanding the contribution of different training phases
- Educational applications: Learning about instruction-following model development
- Development: As a starting point for further preference tuning experiments
For production use cases, we recommend using
Poro 2 70B Instruct instead.
Ethical Considerations and Limitations
Poro 2 70B SFT is an advanced language model optimized for English and Finnish instruction following. As this model has not undergone preference tuning, it may be more prone to generating responses that are misaligned with user expectations compared to the final Instruct model.
Key limitations:
- Limited proficiency in languages other than English and Finnish
- Potential for generating biased or inappropriate content
- May produce factually incorrect information
License
Built with Llama
Poro 2 70B SFT is released under the Llama 3.3 Community License. Please review the license terms before use.
Citation
1@misc{poro2_2025,
2 title={Poro 2: Continued Pretraining for Language Acquisition},
3 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},
4 year={2025},
5 howpublished={LumiOpen}
6}
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.
