|
--- |
|
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 |
|
--- |
|
|
|
## *Model updated on September 24* |
|
|
|
|
|
# Model Card for EuroLLM-1.7B-Instruct |
|
|
|
|
|
This is the model card for the first instruction tuned model of the EuroLLM series: EuroLLM-1.7B-Instruct. You can also check the pre-trained version: [EuroLLM-1.7B](https://huggingface.co/utter-project/EuroLLM-1.7B). |
|
|
|
- **Developed by:** Unbabel, Instituto Superior Técnico, 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 instruction tuned 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: |
|
| | | |
|
|--------------------------------------|----------------------| |
|
| Sequence Length | 4,096 | |
|
| Number of Layers | 24 | |
|
| Embedding Size | 2,048 | |
|
| FFN Hidden Size | 5,632 | |
|
| Number of Heads | 16 | |
|
| Number of KV Heads (GQA) | 8 | |
|
| Activation Function | SwiGLU | |
|
| Position Encodings | RoPE (\Theta=10,000) | |
|
| Layer Norm | RMSNorm | |
|
| Tied Embeddings | No | |
|
| Embedding Parameters | 0.262B | |
|
| LM Head Parameters | 0.262B | |
|
| Non-embedding Parameters | 1.133B | |
|
| Total Parameters | 1.657B | |
|
|
|
## Run the model |
|
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
model_id = "utter-project/EuroLLM-1.7B-Instruct" |
|
tokenizer = AutoTokenizer.from_pretrained(model_id) |
|
model = AutoModelForCausalLM.from_pretrained(model_id) |
|
|
|
text = '<|im_start|>system\n<|im_end|>\n<|im_start|>user\nTranslate the following English source text to Portuguese:\nEnglish: I am a language model for european languages. \nPortuguese: <|im_end|>\n<|im_start|>assistant\n' |
|
|
|
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](https://huggingface.co/google/gemma-2b) and [Gemma-7B](https://huggingface.co/google/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 |
|
| Model | AVG | AVG en-xx | AVG xx-en | en-ar | en-bg | en-ca | en-cs | en-da | en-de | en-el | en-es-latam | en-et | en-fi | en-fr | en-ga | en-gl | en-hi | en-hr | en-hu | en-it | en-ja | en-ko | en-lt | en-lv | en-mt | en-nl | en-no | en-pl | en-pt-br | en-ro | en-ru | en-sk | en-sl | en-sv | en-tr | en-uk | en-zh-cn | ar-en | bg-en | ca-en | cs-en | da-en | de-en | el-en | es-latam-en | et-en | fi-en | fr-en | ga-en | gl-en | hi-en | hr-en | hu-en | it-en | ja-en | ko-en | lt-en | lv-en | mt-en | nl-en | no-en | pl-en | pt-br-en | ro-en | ru-en | sk-en | sl-en | sv-en | tr-en | uk-en | zh-cn-en | |
|
|--------------------------------|------|-----------|-----------|-------|-------|-------|-------|-------|-------|-------|--------------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|----------|-------|-------|-------|-------|-------|-------|-------|----------|-------|-------|-------|-------|-------|-------|-------|--------------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|-------|----------|-------|-------|-------|-------|-------|-------|-------|----------| |
|
| EuroLLM-1.7B-Instruct |86.89 | 86.53 | 87.25 | 85.17 | 89.42 | 84.72 | 89.13 | 89.47 | 86.90 | 87.60 | 86.29 | 88.95 | 89.40 | 87.69 | 74.89 | 86.41 | 76.92 | 84.79 | 86.78 | 88.17 | 89.76 | 87.70 | 87.27 | 87.62 | 67.84 | 87.10 | 90.00 | 88.18 | 89.29 | 89.49 | 88.32 | 88.18 | 86.85 | 90.00 | 87.31 | 87.89 | 86.60 | 86.34 | 87.45 | 87.57 | 87.95 | 89.72 | 88.80 | 87.00 | 86.77 | 88.34 | 89.09 | 88.95 | 82.69 | 87.80 | 88.37 | 86.71 | 87.20 | 87.81 | 86.79 | 86.79 | 85.62 | 86.48 | 81.10 | 86.97 | 90.25 | 85.75 | 89.20 | 88.88 | 86.00 | 87.38 | 86.76 | 89.61 | 87.94 | |
|
| Gemma-2B-EuroBlocks | 81.59 | 78.97 | 84.21 | 76.68 | 82.73 | 83.14 | 81.63 | 84.63 | 83.15 | 79.42 | 84.05 | 72.58 | 79.73 | 84.97 | 40.50 | 82.13 | 67.79 | 80.53 | 78.36 | 84.90 | 87.43 | 82.98 | 72.29 | 68.68 | 58.55 | 83.13 | 86.15 | 82.78 | 86.79 | 83.14 | 84.61 | 78.18 | 75.37 | 80.89 | 78.38 | 84.38 | 84.35 | 83.88 | 85.77 | 86.85 | 86.31 | 88.24 | 88.12 | 84.79 | 84.90 | 82.51 | 86.32 | 88.29 | 54.78 | 86.53 | 85.83 | 85.41 | 85.18 | 86.77 | 85.78 | 84.99 | 81.65 | 81.78 | 67.27 | 85.92 | 89.07 | 84.14 | 88.07 | 87.17 | 85.23 | 85.09 | 83.95 | 87.57 | 84.77 | |
|
| Gemma-7B-EuroBlocks |85.27 | 83.90 | 86.64 | 86.38 | 87.87 | 85.74 | 84.25 | 85.69 | 81.49 | 85.52 | 86.93 | 62.83 | 84.96 | 75.34 | 84.93 | 83.91 | 86.92 | 88.19 | 86.11 | 81.73 | 80.55 | 66.85 | 85.31 | 89.36 | 85.87 | 88.62 | 88.06 | 86.67 | 84.79 | 82.71 | 86.45 | 85.19 | 86.67 | 85.77 | 86.36 | 87.21 | 88.09 | 87.17 | 89.40 | 88.26 | 86.74 | 86.73 | 87.25 | 88.87 | 88.81 | 72.45 | 87.62 | 87.86 | 87.08 | 87.01 | 87.58 | 86.92 | 86.70 | 85.10 | 85.74 | 77.81 | 86.83 | 90.40 | 85.41 | 89.04 | 88.77 | 86.13 | 86.67 | 86.32 | 89.27 | 87.92 | |
|
|
|
|
|
#### WMT-23 |
|
| Model | AVG | AVG en-xx | AVG xx-en | AVG xx-xx | en-de | en-cs | en-uk | en-ru | en-zh-cn | de-en | uk-en | ru-en | zh-cn-en | cs-uk | |
|
|--------------------------------|------|-----------|-----------|-----------|-------|-------|-------|-------|----------|-------|-------|-------|----------|-------| |
|
| EuroLLM-1.7B-Instruct | 82.91 | 83.20 | 81.77 | 86.82 | 81.56 | 85.23 | 81.30 | 82.47 | 83.61 | 85.03 | 84.06 | 85.25 | 81.31 | 78.83 | 79.42 | 86.82 | |
|
| Gemma-2B-EuroBlocks | 79.96 | 79.01 | 80.86 | 81.15 | 76.82 | 76.05 | 77.92 | 78.98 | 81.58 | 82.73 | 82.71 | 83.99 | 80.35 | 78.27 | 78.99 | 81.15 | |
|
| Gemma-7B-EuroBlocks | 82.76 | 82.26 | 82.70 | 85.98 | 81.37 | 82.42 | 81.54 | 82.18 | 82.90 | 83.17 | 84.29 | 85.70 | 82.46 | 79.73 | 81.33 | 85.98 | |
|
|
|
|
|
|
|
#### WMT-24 |
|
| Model | AVG | AVG en-xx | AVG xx-xx | en-de | en-es-latam | en-cs | en-ru | en-uk | en-ja | en-zh-cn | en-hi | cs-uk | ja-zh-cn | |
|
|---------|------|------|-------|----|---|-------|-------|--------|--------|-------|-------|-------|-----| |
|
| EuroLLM-1.7B-Instruct|79.32 | 79.32 | 79.34 | 79.42 | 80.67 | 80.55 | 78.65 | 80.12 | 82.96 | 80.60 | 71.59 | 83.48 | 75.20 | |
|
|Gemma-2B-EuroBlocks| 74.72 | 74.41 | 75.97 | 74.93 | 78.81 | 70.54 | 74.90 | 75.84 | 79.48 | 78.06 | 62.70 | 79.87 | 72.07 | |
|
|Gemma-7B-EuroBlocks| 78.67 | 78.34 | 80.00 | 78.88 | 80.47 | 78.55 | 78.55 | 80.12 | 80.55 | 78.90 | 70.71 | 84.33 | 75.66 | |
|
|
|
|
|
### General Benchmarks |
|
We also compare EuroLLM-1.7B with [TinyLlama-v1.1](https://huggingface.co/TinyLlama/TinyLlama_v1.1) and [Gemma-2B](https://huggingface.co/google/gemma-2b) on 3 general benchmarks: Arc Challenge and Hellaswag. |
|
For the non-english languages we use the [Okapi](https://aclanthology.org/2023.emnlp-demo.28.pdf) 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 |
|
| Model | Average | English | German | Spanish | French | Italian | Portuguese | Chinese | Russian | Dutch | Arabic | Swedish | Hindi | Hungarian | Romanian | Ukrainian | Danish | Catalan | |
|
|--------------------|---------|---------|--------|---------|--------|---------|------------|---------|---------|-------|--------|---------|--------|-----------|----------|-----------|--------|---------| |
|
| EuroLLM-1.7B | 0.3496 | 0.4061 | 0.3464 | 0.3684 | 0.3627 | 0.3738 | 0.3855 | 0.3521 | 0.3208 | 0.3507 | 0.3045 | 0.3605 | 0.2928 | 0.3271 | 0.3488 | 0.3516 | 0.3513 | 0.3396 | |
|
| TinyLlama-v1.1 | 0.2650 | 0.3712 | 0.2524 | 0.2795 | 0.2883 | 0.2652 | 0.2906 | 0.2410 | 0.2669 | 0.2404 | 0.2310 | 0.2687 | 0.2354 | 0.2449 | 0.2476 | 0.2524 | 0.2494 | 0.2796 | |
|
| Gemma-2B | 0.3617 | 0.4846 | 0.3755 | 0.3940 | 0.4080 | 0.3687 | 0.3872 | 0.3726 | 0.3456 | 0.3328 | 0.3122 | 0.3519 | 0.2851 | 0.3039 | 0.3590 | 0.3601 | 0.3565 | 0.3516 | |
|
|
|
|
|
#### Hellaswag |
|
| Model | Average | English | German | Spanish | French | Italian | Portuguese | Russian | Dutch | Arabic | Swedish | Hindi | Hungarian | Romanian | Ukrainian | Danish | Catalan | |
|
|--------------------|---------|---------|--------|---------|--------|---------|------------|---------|--------|--------|---------|--------|-----------|----------|-----------|--------|---------| |
|
| EuroLLM-1.7B | 0.4744 | 0.4760 | 0.6057 | 0.4793 | 0.5337 | 0.5298 | 0.5085 | 0.5224 | 0.4654 | 0.4949 | 0.4104 | 0.4800 | 0.3655 | 0.4097 | 0.4606 | 0.436 | 0.4702 | 0.4445 | |
|
| TinyLlama-v1.1 |0.3674 | 0.6248 | 0.3650 | 0.4137 | 0.4010 | 0.3780 | 0.3892 | 0.3494 | 0.3588 | 0.2880 | 0.3561 | 0.2841 | 0.3073 | 0.3267 | 0.3349 | 0.3408 | 0.3613 | |
|
| Gemma-2B |0.4666 | 0.7165 | 0.4756 | 0.5414 | 0.5180 | 0.4841 | 0.5081 | 0.4664 | 0.4655 | 0.3868 | 0.4383 | 0.3413 | 0.3710 | 0.4316 | 0.4291 | 0.4471 | 0.4448 | |
|
|
|
|
|
## Bias, Risks, and Limitations |
|
|
|
EuroLLM-1.7B-Instruct has not been aligned to human preferences, so the model may generate problematic outputs (e.g., hallucinations, harmful content, or false statements). |
|
|
