Model Description

This model was built by translating the fineweb-edu annotations into 15 languages using a state-of-the-art proprietary LLM for translation, TowerLLM 70B.

The translation model excels at translating entire documents and thus it is the perfect fit to translate the texts we will use to train our classifier.

The classifier is trained for English, German, Spanish, Japanese, Chinese, Russian, Hindi, Czech, Ukrainian, Icelandic, Portuguese, French, Dutch, Italian and Korean. Since it is built on top of mdeberta-v3-base it should be able to generalize across other languages.

Running Model:

To run inference you must install

pip install transformers[torch]
pip install datasets
pip install pandas
pip install tqdm

After installing those libraries you can run the following code:

import pandas as pd
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
from tqdm import tqdm


device = "cuda"
path = "utter-project/EuroFilter-v1"
model = AutoModelForSequenceClassification.from_pretrained(
    path, 
    device_map=device, 
    trust_remote_code=True, 
    torch_dtype=torch.bfloat16
)
tokenizer = AutoTokenizer.from_pretrained(path, use_fast=True)

def get_model_outputs(texts):
    inputs = tokenizer(texts, padding=True, truncation=True, return_tensors="pt", max_length=512).to(model.device)
    with torch.no_grad():
        outputs = model(**inputs)
        score = outputs.logits
        prob = torch.nn.functional.sigmoid(outputs.binary_logits)
    return score.cpu(), prob.cpu()

def batchify_texts(texts, batch_size):
    for i in range(0, len(texts), batch_size):
        yield texts[i:i + batch_size]

# TODO: replace the next line with the texts you want to classify
texts = LIST_WITH_TEXTS_TO_CLASSIFY
batch_size = 64  # Adjust based on your available memory and model capacity
num_batches = (len(texts) + batch_size - 1) // batch_size

all_scores = []
all_probs = []
with tqdm(total=num_batches, dynamic_ncols=True) as pbar:
    for batch_num, batch in enumerate(batchify_texts(texts, batch_size), 1):
        score, probs = get_model_outputs(batch)
        all_scores.append(score)
        all_probs.append(probs)
        pbar.set_description(f"Processing Batch {batch_num}/{num_batches}")
        pbar.update(1)

# SCORES is the output of the regression head and should reflect the
# educational score of the text!
scores = torch.cat(all_scores, dim=0).squeeze()

## BINARY_PRED is the output of the classification head that tells
# if a text has an acceptable educational score or not.
# NOTE: Converting the scores into binary predictions is also possible
all_probs = torch.cat(all_probs, dim=0).squeeze()
binary_pred = (all_probs >= 0.5).numpy().astype(int)

English Results:

When testing the model on an english partition with 37537 samples the results are comparable to the original FineEdu-classifier.

Regression head results:

              precision    recall  f1-score   support

           0       0.80      0.53      0.64      5130
           1       0.80      0.88      0.83     21602
           2       0.63      0.58      0.61      7849
           3       0.54      0.62      0.58      2310
           4       0.62      0.48      0.54       645
           5       0.00      0.00      0.00         1

    accuracy                           0.74     37537
   macro avg       0.56      0.51      0.53     37537
weighted avg       0.74      0.74      0.74     37537

Binary head results:

              precision    recall  f1-score   support

           0       0.98      0.97      0.98     34581
           1       0.71      0.74      0.73      2956

    accuracy                           0.96     37537
   macro avg       0.85      0.86      0.85     37537
weighted avg       0.96      0.96      0.96     37537

Multilingual Results:

If we evaluate on the same texts translated into 15 different languages are almost identical!

Regression head results:

              precision    recall  f1-score   support

           0       0.80      0.50      0.61      5130
           1       0.79      0.87      0.83     21602
           2       0.61      0.58      0.59      7849
           3       0.52      0.61      0.56      2310
           4       0.61      0.38      0.47       645
           5       0.00      0.00      0.00         1

    accuracy                           0.73     37537
   macro avg       0.55      0.49      0.51     37537
weighted avg       0.73      0.73      0.73     37537

Binary head results:

              precision    recall  f1-score   support

           0       0.98      0.97      0.97     34581
           1       0.70      0.71      0.71      2956

    accuracy                           0.95     37537
   macro avg       0.84      0.84      0.84     37537
weighted avg       0.95      0.95      0.95     37537

Citation

If you use our work, please cite:

@misc{martins2025eurollm9B,
      title={EuroLLM-9B: Technical Report}, 
      author={Pedro Henrique Martins and João Alves and Patrick Fernandes and and Nuno M. Guerreiro and Ricardo Rei and Amin Farajian and Mateusz Klimaszewski and Duarte M. Alves and José Pombal and Manuel Faysse and Pierre Colombo and François Yvon and Barry Haddow and José G. C. de Souza and Alexandra Birch and André F. T. Martins},
      year={2025},
      archivePrefix={arXiv},
      primaryClass={cs.CL}, 
}

Funding Information

This is an output of the European Project UTTER (Unified Transcription and Translation for Extended Reality) funded by European Union’s Horizon Europe Research and Innovation programme under grant agreement number 101070631.

For more information please visit https://he-utter.eu/