Instructions to use LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview")
model = AutoModelForCausalLM.from_pretrained("LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview

SGLang

How to use LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview with Docker Model Runner:
```
docker model run hf.co/LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview
```

Poro 2 8B Math Reasoning RL Preview Model Card

Poro 2 8B Math Reasoning RL Preview is a specialized model focused on mathematical reasoning and problem-solving. This preview model was created through reinforcement learning (RL) on top of the Math Reasoning SFT checkpoint. This model excels at mathematical reasoning tasks but is not optimized for general conversational use or other domains.

The model produces reasoning traces in both English and Finnish, adapting the language of its reasoning based on the language of the input prompt. The RL training used separate environments for English and Finnish math problems, a novel approach that maintains strong bilingual reasoning performance.

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 and TensorWave AMD MI325X clusters, 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 Long Model Family

The Poro 2 Long model family extends the Poro 2 models with longer context support and checkpoints trained especially on math reasoning. There are four checkpoints released: a base model, an instruction-tuned model, a math reasoning SFT checkpoint, and the final math reasoning RL checkpoint.

Base Model	Instruct	Math Reasoning SFT	Math Reasoning RL
Poro 2 Long Base	Poro 2 Long Instruct	Poro 2 Long Math Reasoning SFT	Poro 2 Long Math Reasoning RL

Model Overview

Poro 2 8B Math Reasoning RL Preview is based on the Llama 3.1 8B architecture and has been fine-tuned for mathematical reasoning through supervised fine-tuning followed by reinforcement learning.

Hyperparameter	Value
n_parameters	8.03B
n_layers	32
n_heads	32
n_kv_heads	8
d_model	4096
vocab_size	128384
max_sequence_length	131072
base_model	Poro 2 Long Base

Training Process

Context Extension

The base model underwent context extension using LongRoPE scaling to support 131k tokens.

Supervised Fine-Tuning (SFT)

The SFT phase focused on mathematical reasoning and used:

2.2M English math reasoning samples from nvidia/Llama-Nemotron-Post-Training-Dataset
1.4M Finnish translations of the reasoning samples

Reinforcement Learning (RL)

Following SFT, the model was further trained using reinforcement learning with separate English and Finnish math environments:

Dual-environment RL training with language-specific math problem sets
900 RL training steps

Training Dataset

Stage	Model	Dataset	License
CPT	Llama-Poro-2-Long-Base	https://huggingface.co/datasets/HuggingFaceFW/finepdfs-edu	ODC-BY-1.0
CPT	Llama-Poro-2-Long-Base	https://huggingface.co/datasets/HuggingFaceTB/finemath	ODC-BY-1.0
CPT	Llama-Poro-2-Long-Base	https://huggingface.co/datasets/bigcode/starcoderdata	Refer source materials
Instruct SFT	Llama-Poro-2-Long-Instruct	https://huggingface.co/datasets/LumiOpen/poro2-instruction-collection	Llama 3 Community License
Math SFT	Llama-Poro-2-Long-Math-Reasoning-SFT-Preview	https://huggingface.co/datasets/nvidia/Llama-Nemotron-Post-Training-Dataset	CC-BY-4.0
Math SFT	Llama-Poro-2-Long-Math-Reasoning-SFT-Preview	https://huggingface.co/datasets/LumiOpen/Llama-Nemotron-Post-Training-Dataset-SFT-math-FI	CC-BY-4.0
Math RL	Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview	https://huggingface.co/datasets/Kwai-Klear/KlearReasoner-MathSub-30K	Apache-2.0

Post-Training Hyperparameters

SFT

Hyperparameter	Value
Epochs	3
Global batch size	128
Learning rate	5e-5
Max sequence length	32,768
Packing	Yes

Evaluation Results

Poro 2 8B Math Reasoning RL Preview demonstrates strong mathematical reasoning capabilities in both English and Finnish, showing substantial improvements over both the SFT checkpoint and the original Poro 2 8B Instruct model. On AIME 2025, the model achieves performance comparable to models 6x its size.

Model	Math-500 EN	Math-500 FI	AIME25 EN	AIME25 FI	AIME26 EN	AIME26 FI
Poro 2 8B Math Reasoning RL Preview	0.97	0.95	0.607	0.521	0.678	0.572
Poro 2 8B Math Reasoning SFT Preview	0.96	0.928	0.499	0.403	0.604	0.478
Poro 2 8B Instruct	-	-	0.009	0.003	0.010	0.006

All AIME results are avg@64 with temperature 1.0.

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview"
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": "Kuinka monta positiivista kokonaislukujakajaa luvulla 196 on?"}
]

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

outputs = model.generate(
    inputs,
    max_new_tokens=32000,
    temperature=0.7,
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id
)

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

Intended Use

Poro 2 8B Math Reasoning RL Preview is designed for:

Mathematical problem-solving and reasoning tasks
Step-by-step mathematical reasoning with explicit reasoning traces
Math education and tutoring applications
Competition mathematics (e.g., AIME-level problems)
Mathematical reasoning in both English and Finnish

Note: This model is optimized specifically for mathematical reasoning and is not suitable for general conversational AI or other domains.

Ethical Considerations and Limitations

Poro 2 8B Math Reasoning RL Preview is a specialized model optimized for mathematical reasoning in English and Finnish. As a preview release, this model has specific limitations:

Key limitations:

Optimized exclusively for mathematical reasoning; not suitable for general conversation, coding, creative writing, or other domains
This is an early preview release
Reasoning traces are in English or Finnish depending on prompt language; limited proficiency in other languages
May occasionally generate biased, inappropriate, or factually incorrect content

Safety Considerations:

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 Math Reasoning RL Preview is released under the Llama 3.1 Community License. Please review the license terms before use.

Citation

@misc{poro2_long_2025, title={}, author={}, year={}, howpublished={} }

Acknowledgments

We thank CSC - IT Center for Science, Finland for providing access to the LUMI supercomputer, and TensorWave for providing access to AMD MI325X GPU clusters. 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.

Downloads last month: 22

Safetensors

Model size

8B params

Tensor type

BF16

Dataset used to train LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview

Collection including LumiOpen/Llama-Poro-2-8B-Long-Math-Reasoning-RL-Preview

Poro 2 Long

Collection

5 items • Updated 9 days ago