Model updated on September 24

Model Card for EuroLLM-1.7B

This is the model card for the first pre-trained model of the EuroLLM series: EuroLLM-1.7B. You can also check the instruction tuned version: EuroLLM-1.7B-Instruct.

• 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 1.7B parameter multilingual transfomer LLM.
• 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. EuroLLM-1.7B is a 1.7B parameter model trained on 4 trillion tokens divided across the considered languages and several data sources: Web data, parallel data (en-xx and xx-en), and high-quality datasets. EuroLLM-1.7B-Instruct was further instruction tuned on EuroBlocks, an instruction tuning dataset with focus on general instruction-following and machine translation.

Model Description

EuroLLM uses a standard, dense Transformer architecture:

• We use grouped query attention (GQA) with 8 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 256 Nvidia H100 GPUs of the Marenostrum 5 supercomputer, training the model with a constant batch size of 3,072 sequences, which corresponds to approximately 12 million tokens, using the Adam optimizer, and BF16 precision. Here is a summary of the model hyper-parameters:

Run the model

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "utter-project/EuroLLM-1.7B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

text = "English: My name is EuroLLM. Portuguese:"

inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=20)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Results

Machine Translation

We evaluate EuroLLM-1.7B-Instruct on several machine translation benchmarks: FLORES-200, WMT-23, and WMT-24 comparing it with Gemma-2B and Gemma-7B (also instruction tuned on EuroBlocks). The results show that EuroLLM-1.7B is substantially better than Gemma-2B in Machine Translation and competitive with Gemma-7B.

Flores-200

WMT-23

WMT-24

General Benchmarks

We also compare EuroLLM-1.7B with TinyLlama-v1.1 and Gemma-2B on 3 general benchmarks: Arc Challenge and Hellaswag. For the non-english languages we use the Okapi datasets. Results show that EuroLLM-1.7B is superior to TinyLlama-v1.1 and similar to Gemma-2B on Hellaswag but worse on Arc Challenge. This can be due to the lower number of parameters of EuroLLM-1.7B (1.133B non-embedding parameters against 1.981B).

Arc Challenge

Hellaswag

Bias, Risks, and Limitations

EuroLLM-1.7B has not been aligned to human preferences, so the model may generate problematic outputs (e.g., hallucinations, harmful content, or false statements).

Paper

Paper: EuroLLM: Multilingual Language Models for Europe