Djamé Seddah

Wissam Antoun, Virginie Mouilleron, Benoît Sagot et al.

Recent advances in natural language processing (NLP) have led to the development of large language models (LLMs) such as ChatGPT. This paper proposes a methodology for developing and evaluating ChatGPT detectors for French text, with a focus on investigating their robustness on out-of-domain data and against common attack schemes. The proposed method involves translating an English dataset into French and training a classifier on the translated data. Results show that the detectors can effectively detect ChatGPT-generated text, with a degree of robustness against basic attack techniques in in-domain settings. However, vulnerabilities are evident in out-of-domain contexts, highlighting the challenge of detecting adversarial text. The study emphasizes caution when applying in-domain testing results to a wider variety of content. We provide our translated datasets and models as open-source resources. https://gitlab.inria.fr/wantoun/robust-chatgpt-detection

Wissam Antoun, Benoît Sagot, Djamé Seddah

The widespread use of Large Language Models (LLMs), celebrated for their ability to generate human-like text, has raised concerns about misinformation and ethical implications. Addressing these concerns necessitates the development of robust methods to detect and attribute text generated by LLMs. This paper investigates "Cross-Model Detection," by evaluating whether a classifier trained to distinguish between source LLM-generated and human-written text can also detect text from a target LLM without further training. The study comprehensively explores various LLM sizes and families, and assesses the impact of conversational fine-tuning techniques, quantization, and watermarking on classifier generalization. The research also explores Model Attribution, encompassing source model identification, model family, and model size classification, in addition to quantization and watermarking detection. Our results reveal several key findings: a clear inverse relationship between classifier effectiveness and model size, with larger LLMs being more challenging to detect, especially when the classifier is trained on data from smaller models. Training on data from similarly sized LLMs can improve detection performance from larger models but may lead to decreased performance when dealing with smaller models. Additionally, model attribution experiments show promising results in identifying source models and model families, highlighting detectable signatures in LLM-generated text, with particularly remarkable outcomes in watermarking detection, while no detectable signatures of quantization were observed. Overall, our study contributes valuable insights into the interplay of model size, family, and training data in LLM detection and attribution.

Wissam Antoun, Benoît Sagot, Djamé Seddah

Recent advances in NLP have significantly improved the performance of language models on a variety of tasks. While these advances are largely driven by the availability of large amounts of data and computational power, they also benefit from the development of better training methods and architectures. In this paper, we introduce CamemBERTa, a French DeBERTa model that builds upon the DeBERTaV3 architecture and training objective. We evaluate our model's performance on a variety of French downstream tasks and datasets, including question answering, part-of-speech tagging, dependency parsing, named entity recognition, and the FLUE benchmark, and compare against CamemBERT, the state-of-the-art monolingual model for French. Our results show that, given the same amount of training tokens, our model outperforms BERT-based models trained with MLM on most tasks. Furthermore, our new model reaches similar or superior performance on downstream tasks compared to CamemBERT, despite being trained on only 30% of its total number of input tokens. In addition to our experimental results, we also publicly release the weights and code implementation of CamemBERTa, making it the first publicly available DeBERTaV3 model outside of the original paper and the first openly available implementation of a DeBERTaV3 training objective. https://gitlab.inria.fr/almanach/CamemBERTa

Syrielle Montariol, Arij Riabi, Djamé Seddah

Zero-shot cross-lingual transfer learning has been shown to be highly challenging for tasks involving a lot of linguistic specificities or when a cultural gap is present between languages, such as in hate speech detection. In this paper, we highlight this limitation for hate speech detection in several domains and languages using strict experimental settings. Then, we propose to train on multilingual auxiliary tasks -- sentiment analysis, named entity recognition, and tasks relying on syntactic information -- to improve zero-shot transfer of hate speech detection models across languages. We show how hate speech detection models benefit from a cross-lingual knowledge proxy brought by auxiliary tasks fine-tuning and highlight these tasks' positive impact on bridging the hate speech linguistic and cultural gap between languages.

Ghazi Felhi, Joseph Le Roux, Djamé Seddah

Linking neural representations to linguistic factors is crucial in order to build and analyze NLP models interpretable by humans. Among these factors, syntactic roles (e.g. subjects, direct objects,$\dots$) and their realizations are essential markers since they can be understood as a decomposition of predicative structures and thus the meaning of sentences. Starting from a deep probabilistic generative model with attention, we measure the interaction between latent variables and realizations of syntactic roles and show that it is possible to obtain, without supervision, representations of sentences where different syntactic roles correspond to clearly identified different latent variables. The probabilistic model we propose is an Attention-Driven Variational Autoencoder (ADVAE). Drawing inspiration from Transformer-based machine translation models, ADVAEs enable the analysis of the interactions between latent variables and input tokens through attention. We also develop an evaluation protocol to measure disentanglement with regard to the realizations of syntactic roles. This protocol is based on attention maxima for the encoder and on latent variable perturbations for the decoder. Our experiments on raw English text from the SNLI dataset show that $\textit{i)}$ disentanglement of syntactic roles can be induced without supervision, $\textit{ii)}$ ADVAE separates syntactic roles better than classical sequence VAEs and Transformer VAEs, $\textit{iii)}$ realizations of syntactic roles can be separately modified in sentences by mere intervention on the associated latent variables. Our work constitutes a first step towards unsupervised controllable content generation. The code for our work is publicly available.

Ghazi Felhi, Joseph Le Roux, Djamé Seddah

We propose a generative model for text generation, which exhibits disentangled latent representations of syntax and semantics. Contrary to previous work, this model does not need syntactic information such as constituency parses, or semantic information such as paraphrase pairs. Our model relies solely on the inductive bias found in attention-based architectures such as Transformers. In the attention of Transformers, keys handle information selection while values specify what information is conveyed. Our model, dubbed QKVAE, uses Attention in its decoder to read latent variables where one latent variable infers keys while another infers values. We run experiments on latent representations and experiments on syntax/semantics transfer which show that QKVAE displays clear signs of disentangled syntax and semantics. We also show that our model displays competitive syntax transfer capabilities when compared to supervised models and that comparable supervised models need a fairly large amount of data (more than 50K samples) to outperform it on both syntactic and semantic transfer. The code for our experiments is publicly available.

Wissam Antoun, Francis Kulumba, Rian Touchent et al.

French language models, such as CamemBERT, have been widely adopted across industries for natural language processing (NLP) tasks, with models like CamemBERT seeing over 4 million downloads per month. However, these models face challenges due to temporal concept drift, where outdated training data leads to a decline in performance, especially when encountering new topics and terminology. This issue emphasizes the need for updated models that reflect current linguistic trends. In this paper, we introduce two new versions of the CamemBERT base model-CamemBERTav2 and CamemBERTv2-designed to address these challenges. CamemBERTav2 is based on the DeBERTaV3 architecture and makes use of the Replaced Token Detection (RTD) objective for better contextual understanding, while CamemBERTv2 is built on RoBERTa, which uses the Masked Language Modeling (MLM) objective. Both models are trained on a significantly larger and more recent dataset with longer context length and an updated tokenizer that enhances tokenization performance for French. We evaluate the performance of these models on both general-domain NLP tasks and domain-specific applications, such as medical field tasks, demonstrating their versatility and effectiveness across a range of use cases. Our results show that these updated models vastly outperform their predecessors, making them valuable tools for modern NLP systems. All our new models, as well as intermediate checkpoints, are made openly available on Huggingface.

Wissam Antoun, Benoît Sagot, Djamé Seddah

Pretrained transformer-encoder models like DeBERTaV3 and ModernBERT introduce architectural advancements aimed at improving efficiency and performance. Although the authors of ModernBERT report improved performance over DeBERTaV3 on several benchmarks, the lack of disclosed training data and the absence of comparisons using a shared dataset make it difficult to determine whether these gains are due to architectural improvements or differences in training data. In this work, we conduct a controlled study by pretraining ModernBERT on the same dataset as CamemBERTaV2, a DeBERTaV3 French model, isolating the effect of model design. Our results show that the previous model generation remains superior in sample efficiency and overall benchmark performance, with ModernBERT's primary advantage being its support for long context, faster training, and inference speed. However, the new proposed model still provides meaningful architectural improvements compared to earlier models such as BERT and RoBERTa. Additionally, we observe that high-quality pre-training data accelerates convergence but does not significantly improve final performance, suggesting potential benchmark saturation. These findings show the importance of disentangling pretraining data from architectural innovations when evaluating transformer models.

Javier A. Lopetegui, Arij Riabi, Djamé Seddah

Variations in languages across geographic regions or cultures are crucial to address to avoid biases in NLP systems designed for culturally sensitive tasks, such as hate speech detection or dialog with conversational agents. In languages such as Spanish, where varieties can significantly overlap, many examples can be valid across them, which we refer to as common examples. Ignoring these examples may cause misclassifications, reducing model accuracy and fairness. Therefore, accounting for these common examples is essential to improve the robustness and representativeness of NLP systems trained on such data. In this work, we address this problem in the context of Spanish varieties. We use training dynamics to automatically detect common examples or errors in existing Spanish datasets. We demonstrate the efficacy of using predicted label confidence for our Datamaps \cite{swayamdipta-etal-2020-dataset} implementation for the identification of hard-to-classify examples, especially common examples, enhancing model performance in variety identification tasks. Additionally, we introduce a Cuban Spanish Variety Identification dataset with common examples annotations developed to facilitate more accurate detection of Cuban and Caribbean Spanish varieties. To our knowledge, this is the first dataset focused on identifying the Cuban, or any other Caribbean, Spanish variety.

Arij Riabi, Virginie Mouilleron, Menel Mahamdi et al.

The proliferation of radical content on online platforms poses significant risks, including inciting violence and spreading extremist ideologies. Despite ongoing research, existing datasets and models often fail to address the complexities of multilingual and diverse data. To bridge this gap, we introduce a publicly available multilingual dataset annotated with radicalization levels, calls for action, and named entities in English, French, and Arabic. This dataset is pseudonymized to protect individual privacy while preserving contextual information. Beyond presenting our freely available dataset, we analyze the annotation process, highlighting biases and disagreements among annotators and their implications for model performance. Additionally, we use synthetic data to investigate the influence of socio-demographic traits on annotation patterns and model predictions. Our work offers a comprehensive examination of the challenges and opportunities in building robust datasets for radical content detection, emphasizing the importance of fairness and transparency in model development.

Z. Melce Hüsünbeyi, Virginie Mouilleron, Leonie Uhling et al.

The rapid proliferation of misinformation across online platforms underscores the urgent need for robust, up-to-date, explainable, and multilingual fact-checking resources. However, existing datasets are limited in scope, often lacking multimodal evidence, structured annotations, and detailed links between claims, evidence, and verdicts. This paper introduces a comprehensive data collection and processing pipeline that constructs multimodal fact-checking datasets in French and German languages by aggregating ClaimReview feeds, scraping full debunking articles, normalizing heterogeneous claim verdicts, and enriching them with structured metadata and aligned visual content. We used state-of-the-art large language models (LLMs) and multimodal LLMs for (i) evidence extraction under predefined evidence categories and (ii) justification generation that links evidence to verdicts. Evaluation with G-Eval and human assessment demonstrates that our pipeline enables fine-grained comparison of fact-checking practices across different organizations or media markets, facilitates the development of more interpretable and evidence-grounded fact-checking models, and lays the groundwork for future research on multilingual, multimodal misinformation verification.

Nathan Godey, Wissam Antoun, Rian Touchent et al.

We release Gaperon, a fully open suite of French-English-coding language models designed to advance transparency and reproducibility in large-scale model training. The Gaperon family includes 1.5B, 8B, and 24B parameter models trained on 2-4 trillion tokens, released with all elements of the training pipeline: French and English datasets filtered with a neural quality classifier, an efficient data curation and training framework, and hundreds of intermediate checkpoints. Through this work, we study how data filtering and contamination interact to shape both benchmark and generative performance. We find that filtering for linguistic quality enhances text fluency and coherence but yields subpar benchmark results, and that late deliberate contamination -- continuing training on data mixes that include test sets -- recovers competitive scores while only reasonably harming generation quality. We discuss how usual neural filtering can unintentionally amplify benchmark leakage. To support further research, we also introduce harmless data poisoning during pretraining, providing a realistic testbed for safety studies. By openly releasing all models, datasets, code, and checkpoints, Gaperon establishes a reproducible foundation for exploring the trade-offs between data curation, evaluation, safety, and openness in multilingual language model development.

Arij Riabi, Menel Mahamdi, Virginie Mouilleron et al.

Protecting privacy is essential when sharing data, particularly in the case of an online radicalization dataset that may contain personal information. In this paper, we explore the balance between preserving data usefulness and ensuring robust privacy safeguards, since regulations like the European GDPR shape how personal information must be handled. We share our method for manually pseudonymizing a multilingual radicalization dataset, ensuring performance comparable to the original data. Furthermore, we highlight the importance of establishing comprehensive guidelines for processing sensitive NLP data by sharing our complete pseudonymization process, our guidelines, the challenges we encountered as well as the resulting dataset.

Hila Gonen, Ganesh Jawahar, Djamé Seddah et al.

The problem of comparing two bodies of text and searching for words that differ in their usage between them arises often in digital humanities and computational social science. This is commonly approached by training word embeddings on each corpus, aligning the vector spaces, and looking for words whose cosine distance in the aligned space is large. However, these methods often require extensive filtering of the vocabulary to perform well, and - as we show in this work - result in unstable, and hence less reliable, results. We propose an alternative approach that does not use vector space alignment, and instead considers the neighbors of each word. The method is simple, interpretable and stable. We demonstrate its effectiveness in 9 different setups, considering different corpus splitting criteria (age, gender and profession of tweet authors, time of tweet) and different languages (English, French and Hebrew).

Arij Riabi, Benoît Sagot, Djamé Seddah

Recent impressive improvements in NLP, largely based on the success of contextual neural language models, have been mostly demonstrated on at most a couple dozen high-resource languages. Building language models and, more generally, NLP systems for non-standardized and low-resource languages remains a challenging task. In this work, we focus on North-African colloquial dialectal Arabic written using an extension of the Latin script, called NArabizi, found mostly on social media and messaging communication. In this low-resource scenario with data displaying a high level of variability, we compare the downstream performance of a character-based language model on part-of-speech tagging and dependency parsing to that of monolingual and multilingual models. We show that a character-based model trained on only 99k sentences of NArabizi and fined-tuned on a small treebank of this language leads to performance close to those obtained with the same architecture pre-trained on large multilingual and monolingual models. Confirming these results a on much larger data set of noisy French user-generated content, we argue that such character-based language models can be an asset for NLP in low-resource and high language variability set-tings.

José Carlos Rosales Núñez, Guillaume Wisniewski, Djamé Seddah

This work explores the capacities of character-based Neural Machine Translation to translate noisy User-Generated Content (UGC) with a strong focus on exploring the limits of such approaches to handle productive UGC phenomena, which almost by definition, cannot be seen at training time. Within a strict zero-shot scenario, we first study the detrimental impact on translation performance of various user-generated content phenomena on a small annotated dataset we developed, and then show that such models are indeed incapable of handling unknown letters, which leads to catastrophic translation failure once such characters are encountered. We further confirm this behavior with a simple, yet insightful, copy task experiment and highlight the importance of reducing the vocabulary size hyper-parameter to increase the robustness of character-based models for machine translation.

José Carlos Rosales Núñez, Djamé Seddah, Guillaume Wisniewski

This work takes a critical look at the evaluation of user-generated content automatic translation, the well-known specificities of which raise many challenges for MT. Our analyses show that measuring the average-case performance using a standard metric on a UGC test set falls far short of giving a reliable image of the UGC translation quality. That is why we introduce a new data set for the evaluation of UGC translation in which UGC specificities have been manually annotated using a fine-grained typology. Using this data set, we conduct several experiments to measure the impact of different kinds of UGC specificities on translation quality, more precisely than previously possible.

Julien Launay, Elena Tommasone, Baptiste Pannier et al.

Access to large pre-trained models of varied architectures, in many different languages, is central to the democratization of NLP. We introduce PAGnol, a collection of French GPT models. Using scaling laws, we efficiently train PAGnol-XL (1.5B parameters) with the same computational budget as CamemBERT, a model 13 times smaller. PAGnol-XL is the largest model trained to date for the French language. We plan to train increasingly large and performing versions of PAGnol, exploring the capabilities of French extreme-scale models. For this first release, we focus on the pre-training and scaling calculations underlining PAGnol. We fit a scaling law for compute for the French language, and compare it with its English counterpart. We find the pre-training dataset significantly conditions the quality of the outputs, with common datasets such as OSCAR leading to low-quality offensive text. We evaluate our models on discriminative and generative tasks in French, comparing to other state-of-the-art French and multilingual models, and reaching the state of the art in the abstract summarization task. Our research was conducted on the public GENCI Jean Zay supercomputer, and our models up to the Large are made publicly available.

Ghazi Felhi, Joseph Le Roux, Djamé Seddah

Semi-Supervised Variational Autoencoders (SSVAEs) are widely used models for data efficient learning. In this paper, we question the adequacy of the standard design of sequence SSVAEs for the task of text classification as we exhibit two sources of overcomplexity for which we provide simplifications. These simplifications to SSVAEs preserve their theoretical soundness while providing a number of practical advantages in the semi-supervised setup where the result of training is a text classifier. These simplifications are the removal of (i) the Kullback-Liebler divergence from its objective and (ii) the fully unobserved latent variable from its probabilistic model. These changes relieve users from choosing a prior for their latent variables, make the model smaller and faster, and allow for a better flow of information into the latent variables. We compare the simplified versions to standard SSVAEs on 4 text classification tasks. On top of the above-mentioned simplification, experiments show a speed-up of 26%, while keeping equivalent classification scores. The code to reproduce our experiments is public.

Benjamin Muller, Yanai Elazar, Benoît Sagot et al.

Multilingual pretrained language models have demonstrated remarkable zero-shot cross-lingual transfer capabilities. Such transfer emerges by fine-tuning on a task of interest in one language and evaluating on a distinct language, not seen during the fine-tuning. Despite promising results, we still lack a proper understanding of the source of this transfer. Using a novel layer ablation technique and analyses of the model's internal representations, we show that multilingual BERT, a popular multilingual language model, can be viewed as the stacking of two sub-networks: a multilingual encoder followed by a task-specific language-agnostic predictor. While the encoder is crucial for cross-lingual transfer and remains mostly unchanged during fine-tuning, the task predictor has little importance on the transfer and can be reinitialized during fine-tuning. We present extensive experiments with three distinct tasks, seventeen typologically diverse languages and multiple domains to support our hypothesis.

Yves Rychener, Xavier Renard, Djamé Seddah et al.

NLP Interpretability aims to increase trust in model predictions. This makes evaluating interpretability approaches a pressing issue. There are multiple datasets for evaluating NLP Interpretability, but their dependence on human provided ground truths raises questions about their unbiasedness. In this work, we take a different approach and formulate a specific classification task by diverting question-answering datasets. For this custom classification task, the interpretability ground-truth arises directly from the definition of the classification problem. We use this method to propose a benchmark and lay the groundwork for future research in NLP interpretability by evaluating a wide range of current state of the art methods.

Yves Rychener, Xavier Renard, Djamé Seddah et al.

Current methods for Black-Box NLP interpretability, like LIME or SHAP, are based on altering the text to interpret by removing words and modeling the Black-Box response. In this paper, we outline limitations of this approach when using complex BERT-based classifiers: The word-based sampling produces texts that are out-of-distribution for the classifier and further gives rise to a high-dimensional search space, which can't be sufficiently explored when time or computation power is limited. Both of these challenges can be addressed by using segments as elementary building blocks for NLP interpretability. As illustration, we show that the simple choice of sentences greatly improves on both of these challenges. As a consequence, the resulting explainer attains much better fidelity on a benchmark classification task.

Ghazi Felhi, Joseph Le Roux, Djamé Seddah

We present an unsupervised method to obtain disentangled representations of sentences that single out semantic content. Using modified Transformers as building blocks, we train a Variational Autoencoder to translate the sentence to a fixed number of hierarchically structured latent variables. We study the influence of each latent variable in generation on the dependency structure of sentences, and on the predicate structure it yields when passed through an Open Information Extraction model. Our model could separate verbs, subjects, direct objects, and prepositional objects into latent variables we identified. We show that varying the corresponding latent variables results in varying these elements in sentences, and that swapping them between couples of sentences leads to the expected partial semantic swap.

Manuela Sanguinetti, Lauren Cassidy, Cristina Bosco et al.

This article presents a discussion on the main linguistic phenomena which cause difficulties in the analysis of user-generated texts found on the web and in social media, and proposes a set of annotation guidelines for their treatment within the Universal Dependencies (UD) framework of syntactic analysis. Given on the one hand the increasing number of treebanks featuring user-generated content, and its somewhat inconsistent treatment in these resources on the other, the aim of this article is twofold: (1) to provide a condensed, though comprehensive, overview of such treebanks -- based on available literature -- along with their main features and a comparative analysis of their annotation criteria, and (2) to propose a set of tentative UD-based annotation guidelines, to promote consistent treatment of the particular phenomena found in these types of texts. The overarching goal of this article is to provide a common framework for researchers interested in developing similar resources in UD, thus promoting cross-linguistic consistency, which is a principle that has always been central to the spirit of UD.

Benjamin Muller, Antonis Anastasopoulos, Benoît Sagot et al.

Transfer learning based on pretraining language models on a large amount of raw data has become a new norm to reach state-of-the-art performance in NLP. Still, it remains unclear how this approach should be applied for unseen languages that are not covered by any available large-scale multilingual language model and for which only a small amount of raw data is generally available. In this work, by comparing multilingual and monolingual models, we show that such models behave in multiple ways on unseen languages. Some languages greatly benefit from transfer learning and behave similarly to closely related high resource languages whereas others apparently do not. Focusing on the latter, we show that this failure to transfer is largely related to the impact of the script used to write such languages. Transliterating those languages improves very significantly the ability of large-scale multilingual language models on downstream tasks.

Arij Riabi, Thomas Scialom, Rachel Keraron et al.

Coupled with the availability of large scale datasets, deep learning architectures have enabled rapid progress on the Question Answering task. However, most of those datasets are in English, and the performances of state-of-the-art multilingual models are significantly lower when evaluated on non-English data. Due to high data collection costs, it is not realistic to obtain annotated data for each language one desires to support. We propose a method to improve the Cross-lingual Question Answering performance without requiring additional annotated data, leveraging Question Generation models to produce synthetic samples in a cross-lingual fashion. We show that the proposed method allows to significantly outperform the baselines trained on English data only. We report a new state-of-the-art on four multilingual datasets: MLQA, XQuAD, SQuAD-it and PIAF (fr).

Ghazi Felhi, Joseph Leroux, Djamé Seddah

Even though Variational Autoencoders (VAEs) are widely used for semi-supervised learning, the reason why they work remains unclear. In fact, the addition of the unsupervised objective is most often vaguely described as a regularization. The strength of this regularization is controlled by down-weighting the objective on the unlabeled part of the training set. Through an analysis of the objective of semi-supervised VAEs, we observe that they use the posterior of the learned generative model to guide the inference model in learning the partially observed latent variable. We show that given this observation, it is possible to gain finer control on the effect of the unsupervised objective on the training procedure. Using importance weighting, we derive two novel objectives that prioritize either one of the partially observed latent variable, or the unobserved latent variable. Experiments on the IMDB english sentiment analysis dataset and on the AG News topic classification dataset show the improvements brought by our prioritization mechanism and exhibit a behavior that is inline with our description of the inner working of Semi-Supervised VAEs.

Benjamin Muller, Benoit Sagot, Djamé Seddah

Building natural language processing systems for non standardized and low resource languages is a difficult challenge. The recent success of large-scale multilingual pretrained language models provides new modeling tools to tackle this. In this work, we study the ability of multilingual language models to process an unseen dialect. We take user generated North-African Arabic as our case study, a resource-poor dialectal variety of Arabic with frequent code-mixing with French and written in Arabizi, a non-standardized transliteration of Arabic to Latin script. Focusing on two tasks, part-of-speech tagging and dependency parsing, we show in zero-shot and unsupervised adaptation scenarios that multilingual language models are able to transfer to such an unseen dialect, specifically in two extreme cases: (i) across scripts, using Modern Standard Arabic as a source language, and (ii) from a distantly related language, unseen during pretraining, namely Maltese. Our results constitute the first successful transfer experiments on this dialect, paving thus the way for the development of an NLP ecosystem for resource-scarce, non-standardized and highly variable vernacular languages.

Louis Martin, Benjamin Muller, Pedro Javier Ortiz Suárez et al.

Pretrained language models are now ubiquitous in Natural Language Processing. Despite their success, most available models have either been trained on English data or on the concatenation of data in multiple languages. This makes practical use of such models --in all languages except English-- very limited. In this paper, we investigate the feasibility of training monolingual Transformer-based language models for other languages, taking French as an example and evaluating our language models on part-of-speech tagging, dependency parsing, named entity recognition and natural language inference tasks. We show that the use of web crawled data is preferable to the use of Wikipedia data. More surprisingly, we show that a relatively small web crawled dataset (4GB) leads to results that are as good as those obtained using larger datasets (130+GB). Our best performing model CamemBERT reaches or improves the state of the art in all four downstream tasks.