|
--- |
|
license: apache-2.0 |
|
language: |
|
- en |
|
- de |
|
- es |
|
- fr |
|
- it |
|
- pt |
|
- pl |
|
- nl |
|
- tr |
|
- sv |
|
- cs |
|
- el |
|
- hu |
|
- ro |
|
- fi |
|
- uk |
|
- sl |
|
- sk |
|
- da |
|
- lt |
|
- lv |
|
- et |
|
- bg |
|
- 'no' |
|
- ca |
|
- hr |
|
- ga |
|
- mt |
|
- gl |
|
- zh |
|
- ru |
|
- ko |
|
- ja |
|
- ar |
|
- hi |
|
library_name: transformers |
|
--- |
|
|
|
# Model Card for EuroLLM-9B |
|
|
|
|
|
This is the model card for EuroLLM-9B. You can also check the instruction tuned version: [EuroLLM-9B-Instruct](https://huggingface.co/utter-project/EuroLLM-9B-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 9B 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-9B is a 9B 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-9B-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 400 Nvidia H100 GPUs of the Marenostrum 5 supercomputer, training the model with a constant batch size of 2,800 sequences, which corresponds to approximately 12 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 | 42 | |
|
| Embedding Size | 4,096 | |
|
| FFN Hidden Size | 12,288 | |
|
| Number of Heads | 32 | |
|
| Number of KV Heads (GQA) | 8 | |
|
| Activation Function | SwiGLU | |
|
| Position Encodings | RoPE (\Theta=10,000) | |
|
| Layer Norm | RMSNorm | |
|
| Tied Embeddings | No | |
|
| Embedding Parameters | 0.524B | |
|
| LM Head Parameters | 0.524B | |
|
| Non-embedding Parameters | 8.105B | |
|
| Total Parameters | 9.154B | |
|
|
|
## Run the model |
|
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
model_id = "utter-project/EuroLLM-9B" |
|
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 |
|
|
|
### EU Languages |
|
|
|
|
|
 |
|
**Table 1:** Comparison of open-weight LLMs on multilingual benchmarks. The borda count corresponds to the average ranking of the models (see ([Colombo et al., 2022](https://arxiv.org/abs/2202.03799))). For Arc-challenge, Hellaswag, and MMLU we are using Okapi datasets ([Lai et al., 2023](https://aclanthology.org/2023.emnlp-demo.28/)) which include 11 languages. For MMLU-Pro and MUSR we translate the English version with Tower ([Alves et al., 2024](https://arxiv.org/abs/2402.17733)) to 6 EU languages. |
|
\* As there are no public versions of the pre-trained models, we evaluated them using the post-trained versions. |
|
|
|
The results in Table 1 highlight EuroLLM-9B's superior performance on multilingual tasks compared to other European-developed models (as shown by the Borda count of 1.0), as well as its strong competitiveness with non-European models, achieving results comparable to Gemma-2-9B and outperforming the rest on most benchmarks. |
|
|
|
### English |
|
|
|
|
|
 |
|
|
|
**Table 2:** Comparison of open-weight LLMs on English general benchmarks. |
|
\* As there are no public versions of the pre-trained models, we evaluated them using the post-trained versions. |
|
|
|
The results in Table 2 demonstrate EuroLLM's strong performance on English tasks, surpassing most European-developed models and matching the performance of Mistral-7B (obtaining the same Borda count). |
|
|
|
|
|
## Bias, Risks, and Limitations |
|
|
|
EuroLLM-9B has not been aligned to human preferences, so the model may generate problematic outputs (e.g., hallucinations, harmful content, or false statements). |
|
|
