Model Card for EuroMoE-2.6B-A0.6B-Instruct-Preview

⚠️ PREVIEW RELEASE: This is a preview version of EuroMoE-2.6B-A0.6B-Instruct-Preview. The model is still under development and may have limitations in performance and stability. Use with caution in production environments.

This is the model card for EuroMoE-2.6B-A0.6B-Instruct-Preview. You can also check the pre-trained version: EuroMoE-2.6B-A0.6B-Preview.

• Developed by: Unbabel, Instituto Superior Técnico, Instituto de Telecomunicações, University of Edinburgh, Aveni, University of Paris-Saclay, University of Amsterdam, Naver Labs, Sorbonne Université.
• Funded by: European Union.
• Model type: A 2.6B parameter multilingual transformer MoE with 0.6B active parameters.
• Language(s) (NLP): Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Irish, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Arabic, Catalan, Chinese, Galician, Hindi, Japanese, Korean, Norwegian, Russian, Turkish, and Ukrainian.
• License: Apache License 2.0.

Model Details

The EuroLLM project has the goal of creating a suite of LLMs capable of understanding and generating text in all European Union languages as well as some additional relevant languages. EuroMoE-2.6B-A0.6B is a 22B parameter model trained on 8 trillion tokens divided across the considered languages and several data sources: Web data, parallel data (en-xx and xx-en), and high-quality datasets. EuroMoE-2.6B-A0.6B-Instruct was further instruction tuned on EuroBlocks, an instruction tuning dataset with focus on general instruction-following and machine translation.

Model Description

EuroMoE uses a standard MoE Transformer architecture:

• We use grouped query attention (GQA) with 2 key-value heads, since it has been shown to increase speed at inference time while maintaining downstream performance.
• We perform pre-layer normalization, since it improves the training stability, and use the RMSNorm, which is faster.
• We use the SwiGLU activation function, since it has been shown to lead to good results on downstream tasks.
• We use rotary positional embeddings (RoPE) in every layer, since these have been shown to lead to good performances while allowing the extension of the context length.

For pre-training, we use 512 Nvidia A100 GPUs of the Leonardo supercomputer, training the model with a constant batch size of 4096 sequences, which corresponds to approximately 17 million tokens, using the Adam optimizer, and BF16 precision. Here is a summary of the model hyper-parameters:

Sequence Length	4,096
Number of Layers	24
Embedding Size	1,024
Total/Active experts	64/8
Expert Hidden Size	512
Number of Heads	8
Number of KV Heads (GQA)	2
Activation Function	SwiGLU
Position Encodings	RoPE (\Theta=500,000)
Layer Norm	RMSNorm
Tied Embeddings	Yes
Embedding Parameters	0.13B
LM Head Parameters	0.13B
Active Non-embedding Parameters	0.34B
Total Non-embedding Parameters	2.35B
Active Parameters	0.6B
Total Parameters	2.61B

Run the model

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "utter-project/EuroMoE-2.6B-A0.6B-Instruct-Preview"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

messages = [
    {
        "role": "system",
        "content": "You are EuroLLM --- an AI assistant specialized in European languages that provides safe, educational and helpful answers.",
    },
    {
        "role": "user", "content": "What is the capital of Portugal? How would you describe it?"
    },
    ]

inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt")
outputs = model.generate(inputs, max_new_tokens=1024)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Bias, Risks, and Limitations

EuroMoE-2.6B-A0.6B-Instruct-Preview has not been aligned to human preferences, so the model may generate problematic outputs (e.g., hallucinations, harmful content, or false statements).