Universal CEFR: Enabling Open Multilingual Research on Language Proficiency Assessment

Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 9715–9767 November 4-9, 2025 ©2025 Association for Computational Linguistics UNIVERSALCEFR: Enabling Open Multilingual Research on Language Proficiency Assessment Joseph Marvin Imperial1,3, Abdullah Barayan2,14, Regina Stodden4, Rodrigo Wilkens5,Ricardo Muñoz Sánchez6,Lingyun Gao7,Melissa Torgbi1, Dawn Knight2,Gail Forey1,Reka R. Jablonkai1,Ekaterina Kochmar8, Robert Reynolds9,Eugénio Ribeiro10,11,Horacio Saggion12, Elena Volodina6,Sowmya Vajjala13,Thomas François7, Fernando Alva-Manchego2,Harish Tayyar Madabushi1 1University of Bath, 2Cardiff University, 3National University Philippines, 4Bielefeld University, 5University of Exeter, 6University of Gothenburg, 7UCLouvain, 8MBZUAI, 9Brigham Young University, 10INESC-ID Lisboa, 11Instituto Universitário de Lisboa (ISCTE-IUL), 12Universitat Pompeu Fabra, 13National Research Council, Canada, 14King Abdulaziz University CORRESPONDENCE:[email protected],[email protected] Abstract We introduce UNIVERSALCEFR, a largescale multilingual and multidimensional dataset of texts annotated with CEFR (Common European Framework of Reference) levels in 13 languages. To enable open research in automated readability and language proficiency assessment, UNIVERSALCEFR comprises 505,807 CEFR-labeled texts curated from educational and learner-oriented resources, standardized into a unified data format to support consistent processing, analysis, and modelling across tasks and languages. To demonstrate its utility, we conduct benchmarking experiments using three modelling paradigms: a) linguistic feature-based classification, b) fine-tuning pre-trained LLMs, and c) descriptor-based prompting of instruction-tuned LLMs. Our results support using linguistic features and fine-tuning pretrained models in multilingual CEFR level assessment. Overall, UNIVERSALCEFR aims to establish best practices in data distribution for language proficiency research by standardising dataset formats, and promoting their accessibility to the global research community. universalcefr.github.io huggingface.co/UniversalCEFR github.com/UniversalCEFR 1 Introduction Language proficiency research plays a central role in education, and often intersects with advances in linguistics and artificial intelligence (AI). In natural language processing (NLP), language proficiency has been approached through well-established tasks Dialogue English Arabic Czech Dutch Estonian French German Hindi Italian Portuguese Russian Spanish Welsh Sentence Paragraph Document UniversalCEFR Dataset CATEGORY (2) FORMAT (4) ANNOTATION (2) LEVEL COVERAGE (6,+) Manual Computer-Assisted Reference Text LANGUAGES (13) Learner Text A1 A2 B1 B2 C1 C2 + UniversalCEFR Initiative Languages (13) Arabic Czech Dutch English Estonian French German Hindi Italian Portuguese Russian Spanish Welsh Categories (2) Learner Text Reference Text Formats (4) Sentence-Level Paragraph-Level Document-Level Dialogue-Level Levels (6) CEFR-recognized A1 B1 C1 A2 B2 C2 Use Cases Readability Assessment* Feature Analysis* Essay Scoring Text Simplification Corpus Analysis Story Generation Open Permissive Licenses Standardized Format Machine-Readable (HF) Open · Multilingual · Multiformat · Multicategory · Multilevel · Multipurpose The UniversalCEFR Dataset Languages (13) Arabic Czech Dutch English Estonian French German Hindi Italian Portuguese Russian Spanish Welsh Categories (2) Learner Text Reference Text Formats (4) Sentence-Level Paragraph-Level Document-Level Dialogue-Level Levels (6) CEFR-recognized A1 B1 C1 A2 B2 C2 Use Cases Readability Assessment Feature Analysis Essay Scoring Text Simplification Corpus Analysis Story Generation Learner Simulation Style Transfer Accessibility Permissive Licenses Standardised Format Machine-Readable Open · Multilingual · Multiformat · Multicategory · Multilevel · Multipurpose Figure 1: Overview of the contributions of the UNIVERSALCEFR dataset, highlighting its diverse structural coverage—spanning language, format, category, and CEFR level—as well as its accessibility and interoperability for downstream tasks and use cases enabled by permissive licenses and standardized data formats. such as automated readability assessment (ARA) and automated essay scoring (AES). ARA focuses on determining whether a given text matches the expected reading skills of language learners according to their level, whereas AES evaluates the writing skills of the learners as reflected in a text they have written. In this paper, we combine these tasks under the more generic term of language proficiency assessment, as it has varied practical applications in educational assessment and calibration of reading materials for learners (Xia et al.,2016; Harsch,2014;Figueras,2012) as well as for various NLP tasks (see use cases in Figure 1). A widely recognized standard for measuring second language (L2) proficiency is the Common European Framework of Reference for Languages (CEFR),1devel1https://www.coe.int/en/web/common-european-f ramework-reference-languages 9715 Resource # Datasets Indexed # Languages Covered Data Types Data Accessibility Standard Format Geographic Restrictions CEFRLex 7†6 text unrestricted no none Corpora @ UCLouvain 31†9 text, audio, video request per corpus no yes CLARIN L2 Learner Corpora 75†34 text, video request per corpus no yes Learner Language (Språkbanken) 15†13 text, audio request per corpus no yes UNIVERSALCEFR 26 13 text unrestricted yes none Table 1: Comparison of existing language learning and language proficiency dataset collections with UNIVERSALCEFR. † indicates that only a subset of the corresponding resource in that repository contains CEFR labels. Among the five repositories, UNIVERSALCEFR is the only non-geo-locked and standardized collection, allowing seamless, unrestricted use for non-commercial research with proper attribution. oped by the Council of Europe. CEFR offers a language-independent guide for evaluating learners’ abilities in reading, writing, listening, and speaking. It defines a six-level scale (A1, A2, B1, B2, C1, and C2) denoting increasing language competency (North,2014,2007). Recent advances in language proficiency assessment have moved from models relying on handcrafted linguistic features to large language models (LLMs), which achieve high performance across diverse predictive and generative tasks through post-training techniques such as supervised finetuning (Devlin et al.,2019;Vaswani et al.,2017) or instruction tuning (Wei et al.,2022). This form of task generalization enables complex linguistic pattern (e.g., features that make a text complex) modelling within unified frameworks for assessing language proficiency on standardized scales like CEFR. Moreover, they can also be extended to low-resource languages, potentially improving automatic assessment through techniques such as cross-lingual transfer (He and Li,2024;Imperial and Kochmar,2023a,b;Vajjala and Rama,2018). To fully leverage the potential of modern approaches for CEFR-level prediction, researchers require access to high-quality datasets with broad coverage across languages, proficiency levels, and text granularity. However, despite the longstanding use of CEFR in educational and NLP research, there are very limited standardized, machine-readable, and openly accessible collections of CEFR-annotated corpora, especially in terms of language coverage and granularity beyond sentence level (Naous et al.,2024). Moreover, most existing single-language resources are available in inconsistent or outdated formats (e.g., unprocessed text files, XML), which require extensive preprocessing and normalization. Finally, many datasets are restricted by copyright or licensing terms, limiting their accessibility for open research. To this end, our work addresses the resource gap in CEFR-based language proficiency assessment research through the following contributions: • We introduce UNIVERSALCEFR, a largescale multilingual multidimensional open dataset composed of 505K CEFR-labeled texts across 13 languages, designed to advance multilingual research in language proficiency assessment. • We propose a data standardization pipeline and annotation template to homogenize available CEFR-labeled texts, enhancing their interoperability and accessibility for researchers across domains. • We provide a critical reflection of current practices in data sharing of language proficiency assessment resources and suggest pathways towards improvement using UNIVERSALCEFR as a case study for a more open, standardized initiative for resource development. 2 Background Language Learning Databases and Resources. Language learning and language proficiency are research areas driven by the collection of two main types of data: reference-based data created by experts (e.g. reference reading materials) and learner-based data created by language learners (e.g. essays, conversations, and dialogue snippets). If a task requires it, such as in proficiency assessment, these corpora may undergo examination by language proficiency experts who will grade them based on a scale (e.g. CEFR). We list four community-recognized databanks and resource collections in the domain of language learning and proficiency assessment in Table 1. CEFRLex is a collection of machine-readable multilingual 9716 lexicon-based datasets in 6 European languages. The Corpora Hub hosted by UCLouvain, the Learner Language from Språkbanken Text (SBX), and the L2 Learning Corpora hosted by the Common Language Resources and Technology Infrastructure (CLARIN) are all large collections of general multilingual and multimodal language learner datasets. Not all corpora in these databases are annotated with CEFR labels, and each corpus is associated with a publication detailing how they were collected and built and their specific purpose in language learning research. Access Restrictions and Data Privacy Regulations. Despite the existence of L2 resource collections as listed in Table 1, researchers cannot freely and openly use all datasets hosted in these repositories. CEFRLex, 2 Corpora @ UCLouvain, 3 CLARIN, 4 and Språkbanken Text 5 are hosted under European universities and institutions which means they are under the jurisdiction of EU Data Privacy Laws, particularly the General Data Protection Regulation (GDPR). 6 Thus, learner texts from these collections, written based on personal interactions and containing Personally Identifiable Information (PII), can only be accessed through special legal coordination with the data maintainers. If access is granted, the licensee may also need to provide a proof of PII anonymization that produces a derivation distinct from the original dataset as done in Jentoft and Samuel (2023) for the ASK Corpus (Tenfjord et al.,2006) containing L2 Norwegian CEFR-labeled texts and the International Corpus of Learner Finnish (ICLFI) (Jantunen et al., 2013) containing L2 Finnish CEFR-labeled texts. Moreover, some datasets such as the SweLL Corpus (Volodina,2024;Volodina et al.,2019,2016) from Språkbanken Text, composed of Swedish L2 texts with CEFR levels, are geographically licensed and can only be used by institutions within the EU and EEA region. As such, these datasets remain off-limits to any researcher outside of Europe. CEFR Assessment and Standardization. The majority of research on automatic classification (or ranking) of texts based on the CEFR scale 2https://cental.uclouvain.be/cefrlex/ 3https://corpora.uclouvain.be/catalog/ 4https://www.clarin.eu/resource-families/L2-c orpora 5https://spraakbanken.gu.se/en/resources/lear ner-language 6https://gdpr-info.eu/ tends to focus on single-language model evaluations (Ribeiro et al.,2024a;Wilkens et al.,2024, 2023,2018;Tack et al.,2017;Volodina et al., 2016;Pilán et al.,2016;Vajjala and Lõo,2014; Xia et al.,2016;Yancey et al.,2021;VásquezRodríguez et al.,2022). This allows deeper investigation of language-specific nuances and intricacies connected to measuring text complexity. Meanwhile, other works have explored universal, language-agnostic features such as Azpiazu and Pera (2019); Arhiliuc et al. (2020); Caines and Buttery (2020); Vajjala and Rama (2018) where they used traditional word and PoS-ngram features to build a multiand cross-lingual CEFR proficiency classifier for German, Czech, Italian, Spanish, and English, among others. He and Li (2024), on the other hand, focused on cross-lingual automatic essay scoring anchored on the CEFR scale, covering six languages (Czech, English, German, Italian, Portuguese, and Spanish). In parallel with the rise of benchmarking studies for LLMs, similar efforts are growing in the CEFRbased language proficiency community. Two works in this direction include Naous et al. (2024), which introduced ReadMe++, a multilingual, multidomain dataset for sentence-level readability assessment on a CEFR scale covering five languages, while the iRead4Skills Project by Pintard et al. (2024) released a collection of written texts in French, Portuguese, and Spanish across multiple genres and levels patterned to CEFR. Likewise, in data collection standardization, CLARIN released the Core Metadata Schema for Learner Corpora (LC-meta), which aims to provide a structured method with a specific emphasis on capturing metadata of collected learner texts, focusing on learner background, context, and individual differences (Paquot et al.,2024). 3 The UNIVERSALCEFR Dataset To support multilingual language proficiency research, we introduce UNIVERSALCEFR, a largescale initiative that curates and standardizes open human-annotated CEFR-labeled corpora. Unifying diverse resources under a consistent format enables reproducible and scalable research across linguistics, NLP, and education. In this section, we outline the dataset’s design principles, detail the data collection and standardization pipeline, provide key statistics, and present a linguistic feature analysis that supports downstream modelling. 9717 3.1 Design Principles Our methodology was guided by three key design principles. Openness and Accessibility. In building UNIVERSALCEFR, we aim to demonstrate how datadriven research in language proficiency and assessment benefits from standardized, unified data formats. This enables portability and interoperability across domains with evolving data pipelines, such as language model pre-training in NLP. All corpora included in UNIVERSALCEFR are publicly available for non-commercial research through permissive licenses (e.g. Creative Commons). However, significant effort was required to collate and standardize these datasets, highlighting the need for standardization and improved accessibility. Multilinguality and Structure Diversity. Although CEFR originated in Europe, it has been increasingly adopted as a reference framework for language proficiency assessment worldwide. Accordingly, UNIVERSALCEFR extends beyond European languages. Its current version includes 13 languages, spanning high-resource (English, Spanish, French, German, Italian, Portuguese), mid-resource (Dutch, Russian, Arabic), and lowresource (Czech, Estonian, Hindi, Welsh) languages. It also captures structural diversity by annotating each corpus with its production category (learner or reference), granularity (sentence, paragraph, document, or discourse), and label coverage (standard CEFR or CEFR plus levels). Global Collaboration. From its conceptualization and planning, the UNIVERSALCEFR initiative involved close collaboration among 20 researchers in language proficiency assessment, NLP, and education from 13 institutions across nine countries (UK, Canada, USA, Germany, Sweden, UAE, Spain, Belgium, and Portugal). 7 They all played a key role in defining the standardization protocol, designing evaluation experiments, and discussing future research directions. These collaborative decisions are detailed in the following sections. 3.2 Data Collection This section outlines the corpus selection criteria and the standardization methods used in UNIVERSALCEFR for acquiring and consolidating a large and diverse collection of resources. 7 As CEFR is a European framework, most active researchers in the field are based in Europe. Corpora Selection. The inclusion of datasets in UNIVERSALCEFR is guided by three criteria: 1. Public Accessibility: Datasets must be available under a permissive license for noncommercial research (e.g., Creative Commons, CC-BY-NC), or be in the public domain and acquirable through direct download or via a request form for usage tracking. 2. Gold-Standard CEFR Labels: Datasets must include CEFR annotations produced or validated by domain experts, such as language teachers or proficiency researchers, particularly in the case of learner texts. 3. Human Authorship: All texts must be written by humans to ensure suitability for research involving creative, multilingual, multilevel, and multi-genre content. As of this writing, UNIVERSALCEFR does not include machine-generated texts. The full list of consolidated corpora that meet all three UNIVERSALCEFR inclusion criteria is provided in Table 22 in the Appendix. Standardization Process. To ensure interoperability, transformation, and machine readability, we standardized the collected datasets by preprocessing their varied source formats into a unified structure. We adopted JSON as the per-instance format and defined eight metadata fields considered essential for each CEFR-labeled text. These fields include the source dataset, language, granularity (document, paragraph, sentence, discourse), production category (learner or reference), and license. Full descriptions and predetermined values used for each field are provided in Table 15. The final standardized dataset is available from HuggingFace Dataset repository. 8 A key challenge was the lack of a unified format across the language proficiency community. Source corpora came in various formats, including plain text (e.g., csv, tsv, txt), spreadsheets (e.g., XLSX, XLS), markup (e.g., XML), and PDFs requiring manual extraction. This challenge further motivates the need for unified data aggregation initiatives that UNIVERSALCEFR aims to help establish. 8https://huggingface.co/UniversalCEFR/dataset s 9718 UNIVERSALCEFR # of Instances -FULL* 505,807 (out-of-scope instances) 11,316 -FULL 494,491 -TRAIN 435,919 -DEV 54,107 -TEST 4,465 Table 2: Data splits for UNIVERSALCEFR. FULL* denotes all instances, including those with CEFR labels that we currently do not recognize for the task (e.g., NA, A+, B). These were excluded from the TRAIN,DEV, and TEST sets used in our experiments. 3.3 Dataset Statistics The final UNIVERSALCEFR collection comprises 505,807 CEFR-labeled texts across 13 languages and 4 scripts (Latin, Arabic, Devanagari, and Cyrillic). Tables 2and 3show the overall dataset size, its splits and breakdown per CEFR level per language. We identified 11,316 instances with invalid or out-of-scope labels (e.g., NA, A+, B) outside the six recognized CEFR labels (A1–C2) and duplicates, which were removed before splitting UNIVERSALCEFR into TRAIN,DEV, and TEST. For the TEST, we set a cap of 200 instances per language and per granularity level. Additional dataset statistics can be found in Appendix A. LANG A1 A2 B1 B2 C1 C2 EN 192,596 132,614 66,425 23,266 8,004 795 ES 8,282 8,648 6,835 5,061 3,224 0 DE 319 15,970 15,630 474 130 426 NL 51 216 782 738 219 85 CS 1 188 165 81 4 0 IT 29 381 394 2 0 0 FR 151 390 575 478 293 126 ET 0 395 588 407 307 0 PT 314 325 367 233 112 72 AR 81 259 625 645 361 183 HI 263 283 286 263 222 174 RU 402 293 409 326 237 91 CY 764 608 0 0 0 0 Total 203,253 160,570 93,081 31,974 13,113 1,952 Table 3: Data statistics of UNIVERSALCEFR-FULL in terms of recognized CEFR levels (A1, A2, B1, B2, C1, C2) across the 13 target languages. 4 Linguistic Feature Analysis We aim to examine how well a broad set of linguistic features aligns with CEFR proficiency levels across languages in UNIVERSALCEFR. We extracted a set of 100 linguistic features, grouped into morphosyntactic (62), syntactic (18), lengthconcreteness sentence_fre imageability get_type_token_ratio get_ratio_of_punctuation count_characters_per_sentence count_syllables_in_sentence count_characters doc_num_tokens count_words_per_sentence arcscydeenesetfrhiitnlptru -0.21 -0.41 -0.16 -0.31 0.44 0.44 0.40 0.41 -0.67 -0.51 -0.69 -0.52 -0.13 0.62 0.61 0.62 0.73 0.59 -0.06 -0.24 -0.08 -0.17 -0.20 0.32 0.32 0.24 0.29 -0.04 -0.11 -0.02 -0.01 0.04 0.28 0.15 0.15 0.06 0.23 -0.56 -0.11 -0.57 -0.15 -0.48 0.47 0.45 0.47 0.66 0.42 -0.69 -0.27 -0.65 -0.34 -0.45 0.49 0.49 0.49 0.59 0.45 -0.76 -0.57 -0.76 -0.43 -0.40 0.75 0.75 0.75 0.61 0.71 -0.44 -0.61 -0.37 -0.51 -0.07 0.67 0.69 0.68 0.58 0.64 -0.52 -0.70 -0.48 -0.52 -0.22 0.73 0.75 0.70 0.68 -0.59 -0.31 -0.54 -0.52 -0.21 0.60 0.58 0.60 0.75 0.58 -0.18 -0.37 -0.16 -0.27 0.43 0.44 0.43 0.18 0.39 -0.42 -0.37 -0.40 -0.27 -0.18 0.34 0.35 0.34 0.44 0.33 -0.40 -0.59 -0.28 -0.34 -0.13 0.75 0.75 0.75 0.63 0.63 Figure 2: Highly correlated linguistic features occurring in at least three languages. Blue features lean towards positive correlation, while red features denote negative correlation. For brevity, these top features are those lying at the extreme ends of the correlation spectrum. based (11), lexical (4), readability (2), psycholinguistic (2), and discourse (1) categories. A complete and detailed list is available in Appendix E. 4.1 Correlation Across All Languages Considering the absolute Spearman correlation between the features and the CEFR level (selecting values with p < 0.05 and ρ > 0.3 on average across all languages), the strongest associations were found in length-based measures, such as characters per sentence and syllables per sentence. Several grammatical complexity features, including parse tree height and phrase length, showed moderate correlations. Readability indices (FKGL and Flesch Reading Ease) also displayed moderate correlations in the expected direction. Psycholinguistic features, such as concreteness and imageability, were negatively correlated with proficiency, indicating a shift toward more abstract language at higher levels. Finally, morphosyntactic features regarding voice, tense, and number showed moderate but consistent correlations, supporting their relevance in reflecting syntactic development. 9719 4.2 Correlation By CEFR Level To assess the consistency of feature relevance across languages, we examined the number of features with significant correlations ( p < 0.05 ) with CEFR levels per language as visualized in Figure 2. The results revealed notable variations. Languages such as Czech (CS), Estonian (ET), and Italian (IT) showed a high number of relevant features, suggesting strong alignment between the selected linguistic features and CEFR progression in these languages. English (EN), Spanish (ES), French (FR), Hindi (HI), and Russian (RU) showed moderate coverage, with a reasonable number of features exceeding the 0.3 correlation threshold. In contrast, Arabic (AR), Dutch (NL), and Portuguese (PT) exhibited weak coverage, while Welsh (CY) and German (DE) had very few or no features with relevant correlations, indicating a limited match between the current feature set and CEFR levels for those languages. Furthermore, a few features are only relevant for a few languages, e.g., the translative case for only Estonian, negative verb polarity for only Czech, or genitive case for only Czech, Estonian, and Russian. This variability highlights the influence of language-specific properties on the effectiveness of general feature-based models for proficiency prediction. 5 CEFR Level Classification Given the availability of gold-standard CEFR labels and the linguistic diversity of the UNIVERSALCEFR dataset, we define our primary experimental task as multiclass, multilingual CEFR level classification. The goal is to predict one of the six CEFR levels (A1–C2) for a given text instance in any of the 13 supported languages. We evaluate three modeling paradigms: feature-based classification, fine-tuning of multilingual pre-trained models, and prompting LLMs. 5.1 Feature-Based Models We evaluated two widely-used classification models from Scikit-Learn (Pedregosa et al.,2011): Random Forest (RANDFOREST) and Logistic Regression (LOGREGR). Both models were trained on the linguistic features described in Section 4, using Scikit-Learn’s default hyperparameter settings. We experimented with two feature configurations: one using all 100 features (ALLFEATS) and another using an automatically selected subset of top-performing features across all languages (TOPFEATS). Appendices E.1 and E.2 detail the linguistic feature information for both setups. 5.2 Fine-tuned Models We used three BERT-based models with varying degrees of multilingual coverage: ModernBERT (Warner et al.,2024), a monolingual English model with 395M parameters; EuroBERT (Boizard et al., 2025), a multilingual model trained on 15 diverse European and non-European languages, with 210M parameters; and XLM-R (Conneau et al.,2020), a massively multilingual model supporting 100 languages, with 279M parameters. Each model was fine-tuned for three epochs, with the best checkpoint selected based on the highest weighted F1 score on the validation set. Additional details can be found in Appendix Table 17. 5.3 Descriptor-Based Prompting We evaluated three instruction-tuned models: Gemma 1 (Gemma Team,2024), an Englishcentric model with 7B parameters; Gemma 3(Gemma Team,2025), a multilingual model trained on 140+ global languages with 12B parameters; and EuroLLM (Martins et al.,2024), a multilingual model trained on 15 European-centric languages with 9B parameters. We explored five prompting strategies, ranging from no context to setups using CEFR level descriptors for reading comprehension and written production, either in English or in specific languages. The prompt configurations are as follows: • BASE. Generic prompting with no CEFR level descriptors as context. • EN-READ. CEFR level descriptors for reading comprehension in English used as context. • EN-WRITE. CEFR level descriptors for written production in English used as context. • LANG-READ. CEFR level descriptors for reading comprehension, translated to the target language being assessed used as context. • LANG-WRITE. CEFR level descriptors for written production, translated to the target language being assessed used as context. All CEFR descriptors were retrieved from the official CEFR website. Prompt templates and hyperparameter values for each setup are detailed in Table 18 and Appendix I. 5.4 Evaluation Metrics We use weighted F1 as the primary evaluation metric across all experiments. This accounts 9720 MODEL & SETUP EN ES DE NL CS IT FR ET PT AR HI RU CY Avg BASELINE MOST FREQUENT CLASS 7.39 18.1 26.8 21.4 23.8 35.5 16.3 15.9 10.0 23.3 7.28 10.7 33.4 19.3 GEMMA1-7B (ENGLISH) BASE 21.8 26.0 40.6 32.1 44.0 57.3 32.2 39.0 14.0 28.9 25.0 34.8 48.7 34.2 EN-READ 20.5 28.3 31.0 23.5 53.6 41.0 22.7 24.9 27.2 29.5 8.4 18.0 55.7 29.6 EN-WRITE 19.8 24.5 34.5 29.3 51.9 57.7 27.7 42.7 22.2 20.8 14.0 27.6 52.1 32.9 LANG-READ 20.5 29.3 35.1 37.8 55.3 48.0 27.1 44.6 20.2 32.2 12.8 26.2 52.8 34.0 LANG-WRITE 19.8 29.8 32.6 34.0 49.9 61.7 26.3 46.3 21.2 36.7 12.7 26.9 53.6 34.7 GEMMA3-12B (MULTI) BASE 28.8 35.0 42.2 47.0 42.6 65.2 38.1 39.5 24.6 41.8 28.7 29.7 40.9 38.8 EN-READ 19.3 25.5 35.8 25.5 18.5 22.9 29.3 26.0 9.8 33.3 14.8 21.2 20.5 23.3 EN-WRITE 26.6 36.7 46.4 46.7 50.1 77.4 40.5 43.8 27.3 48.6 24.0 37.4 52.4 43.2 LANG-READ 19.3 28.1 35.2 37.6 50.9 64.8 35.0 30.4 26.1 29.5 20.5 32.5 61.6 36.3 LANG-WRITE 26.6 33.2 38.3 39.6 55.0 76.4 37.7 42.4 25.4 38.0 24.6 31.5 53.7 40.2 EUROLLM-9B (MULTI) BASE 18.6 25.4 28.0 29.1 25.0 39.9 25.9 32.0 16.4 34.3 12.7 15.1 14.4 24.4 EN-READ 23.1 26.9 38.1 30.2 33.3 41.9 24.5 33.6 19.9 33.8 18.0 21.8 26.4 28.6 EN-WRITE 21.5 26.2 29.8 32.0 32.4 33.1 26.8 32.8 21.1 31.8 17.7 17.5 24.5 26.7 LANG-READ 23.1 27.0 32.7 31.8 29.8 32.9 28.3 28.6 16.8 32.4 14.3 16.2 17.3 25.5 LANG-WRITE 21.5 28.5 35.1 30.1 30.8 30.6 27.6 29.9 16.5 35.2 21.0 16.1 8.80 25.5 FINE-TUNED MODELS MODERNBERT (ENGLISH) 75.8 71.8 72.1 54.2 66.9 82.7 47.2 88.3 33.5 30.8 51.6 48.9 73.2 61.3 EUROBERT (MULTI) 74.6 72.0 70.6 53.2 63.9 79.7 42.0 86.6 32.1 35.4 44.7 45.9 79.9 60.0 XLM-R (MULTI) 75.5 69.6 73.2 59.0 68.8 83.2 51.6 88.8 29.2 43.0 52.8 49.6 72.6 62.8 FEATURE-BASED MODELS RANDFOREST (TOPFEATS) 62.0 57.6 64.9 54.5 69.5 79.9 44.1 84.2 27.8 43.8 44.1 47.2 72.9 57.9 RANDFOREST (ALLFEATS) 63.4 60.6 65.4 53.0 69.2 79.3 41.4 84.2 26.4 42.8 46.8 47.8 78.2 58.3 LOGREGR (ALLFEATS) 32.1 28.2 50.9 47.1 62.9 81.9 41.7 67.5 23.1 34.1 47.8 41.1 63.8 47.9 LOGREGR (TOPFEATS) 30.4 29.7 52.5 44.1 62.7 82.7 40.3 67.5 22.7 33.5 48.4 41.1 59.2 47.3 Table 4: Full weighted F1 performance results from the multilingual and English-centric model evaluation experiments using three setups (feature-based, fine-tuning, and prompting) and using UNIVERSALCEFR-TEST split across the 13 languages. Boldfaced values indicate the highest scores overall per model setup, while underlined values highlight the highest scores for each model setup within each language. for the class imbalance in CEFR level distribution and granularity across language subsets in UNIVERSALCEFR-TEST. Using accuracy in the experiments would produce misleading performance in favor of any majority class. 6 Results 6.1 Model-Based Performance Comparison Table 4shows that, in terms of overall average performance across languages, the fine-tuned setup with ModernBERT, EuroBERT, and XLMR achieved the highest weighted F1 score range ( ≈ 60%-62.8%) outperforming feature-based models ( ≈ 47%-58%) and prompting ( ≈ 23%-43%). Among the LLM-based approaches—prompting and fine-tuning—models trained on broader multilingual corpora generally performed better. For instance, XLM-R, which supports 100 languages, was the top performer, followed by EuroBERT (15 languages) and ModernBERT (English-only). A similar trend was observed in prompting: Gemma 3, trained on 140+ languages, outperformed EuroLLM (15 languages) and the English-centric Gemma 1, achieving the best prompting score of 43.2. These findings are consistent with previous work (Naous et al.,2024;Shardlow et al.,2024; Colla et al.,2023;Yuan and Strohmaier,2021), reinforcing the usefulness of multilingual models for language proficiency assessment tasks. One limitation of our experimental setup, however, is that we did not include language-specific pre-trained models for languages other than English, which may have further improved performance for low9721 MODEL SENT PARA DOC ALL GEMMA1 19.41 42.74 30.81 33.63 GEMMA3 38.71 43.12 39.62 42.33 XLM-R 62.67 66.38 71.12 65.92 RANDFOREST-ALL 56.88 62.77 64.58 61.38 RANDFOREST-TOP 53.89 62.98 64.94 60.50 Table 5: Weighted F1 scores for top-performing unique model evaluation setups across granularities available for all languages. and mid-resource languages. 6.2 Granularity-Level Comparison Table 5highlights clear performance differences across text granularities (sentence, paragraph, and document) for all models, but more prominently for the Gemma models under prompting. Gemma 1, in particular, tends to over-predict lower CEFR levels (A1–B1) on sentence-level data, whereas its predictions on document-level subsets are more evenly distributed and better aligned with ground truth distributions. This suggests that prompt-based methods may require longer texts to make more accurate predictions, unlike models trained or fine-tuned on the respective datasets. Other models, such as XLM-R and Random Forest, show better results on document ( ≈ 64%-71%) and paragraph-level data ( ≈ 62%-66%) than sentence-level data ( ≈ 53%- 62%), which was shown to be a more difficult task in previous work on readability (Dell’Orletta et al.,2011;Vajjala and Meurers,2014). Regarding language-specific differences, among English, German, and Welsh, the best performance is seen with the paragraph-level dataset for English, the document-level dataset for German, and the sentence-level dataset for Welsh and French with the fine-tuned XLM-R model. Similar variations can be observed for other languages with more than one level of granularity (see Table 19). No single granularity or model shows consistently better performance across all tested languages. These results are likely due to the distribution of excerpts across granularity levels in each language (see Table 7in Appendix A). 6.3 Learner-Reference Comparison Four languages in UNIVERSALCEFR contain both learner and reference texts: Arabic, German, English, and Spanish. Table 6reports the average weighted F1 performance difference between the two categories across the four languages. For GerLANGUAGE LEARNER REFERENCE AR 41.92†54.69 DE 71.14 74.39 EN 83.41 58.24 ES 97.99 42.72 Table 6: Average performances of the best models on learner text versus reference text across languages. † indicates performance with Gemma 3, and the rest refer to performance of the XLM-R model. Only these four languages have both learner and reference texts. man, performance is comparable between learner and reference texts ( ≈ 71–74%). In contrast, English and Spanish show higher performance on learner texts (83% and 98%) than on reference texts (58% and 42%, respectively). Arabic displays the opposite trend: results on reference texts (54%) are much higher than those of learner texts, where the best results were obtained by Gemma 3 (41%). One possible explanation is that Gemma 3 may have been exposed to more Arabic content in its preand post-training phases. 7 Discussion We discuss potential pathways through which UNIVERSALCEFR can serve as a model, and offer key considerations for advancing data accessibility in language proficiency research. Critical Reflections of Current Practices. The multiregional and multidisciplinary effort behind UNIVERSALCEFR exposed significant inconsistencies and critical gaps in building CEFR-labeled language proficiency assessment corpora. Upon examination of annotation practices, there appears to be no standard method for conducting expert annotations, including inconsistent use of interannotator agreement metrics and unclear guidelines on the number of annotators required to achieve reliable agreement. This is reflected in the UNIVERSALCEFR dataset itself, where nearly half of the corpora lack information on the annotators involved and their agreement scores. We posit that this may be due to diverse judgments of what constitutes high-quality data that does not require further human annotations. In terms of language coverage, UNIVERSALCEFR includes nine (EN, ES, DE, NL, CS, IT, FR, ET, PT) of the 24 recognized European languages. As a result, researchers working on these nine languages now have access to open, standard9722 ized data for CEFR-based language proficiency assessment. The remaining 15 languages represent valuable opportunities for future expansion through collaborative efforts. While our open data and standardization initiative is a step towards addressing current challenges in interoperability and accessibility of resources, similar parallel efforts are needed in areas such as annotation and evaluation practices to ensure sustained progress in the language proficiency assessment community. Need for Pro-Research Data Sharing Policies. As generative AI, particularly LLMs, becomes more ubiquitous, organizations that create valuable data for language proficiency assessment, such as publishers, educational institutions, and media outlets, are growing more cautious about how their resources are used. A major concern is the risk of data being used to train proprietary generative models, especially when such models are only accessible via commercial APIs that require transferring evaluation corpora to external servers. An example is the TCFLE-8 corpus (Wilkens et al.,2023) containing CEFR-labeled essays hosted by France Education International. Researchers seeking access to this dataset must explicitly specify that the resource will not be processed through commercial APIs to prevent potential data harvesting. To address these concerns, we believe the community needs to agree on a unified pro-research data sharing policy with clear usage guidelines for academic, non-commercial studies that require analysis of protected data with generative AI models without training on them. Linguistic Features and Fine-tuning Still Matter. While recent advances in LLMs keep transforming NLP research, our multilingual and multidimensional experiments in Section 6reaffirm the continued value of linguistic features for traditional ML classifiers and fine-tuning pre-trained models in language proficiency assessment. We observe common patterns where higher distribution and instance count lead to better results using these two setups (see performances on Spanish, English, and German subsets in Table 4) over prompting with CEFR descriptors. Moreover, using linguistic features in language proficiency assessment allows deeper analysis of language interactions with variables such as complexity, as seen in Appendix C. Given these insights, we encourage further efforts in the expansion of existing but low-resource language datasets with CEFR labels, as well as the exploration of features to better model morphologically-rich languages (e.g., Estonian and Portuguese). Together, these recommendations bridge current observed model failures to practical approaches in improving multilingual CEFR proficiency assessment. 8 Conclusion and Future Directions In this work, we introduced UNIVERSALCEFR, a large-scale, open, multilingual, multidimensional dataset comprising 505,807 CEFR-annotated texts across 13 languages developed through global collaboration. Our findings from diverse model experiments with CEFR level prediction provide strong support for the utility of linguistic features and finetuning multilingual models in language proficiency assessment. Similarly, our critical analysis of the current data and resource-building practices emphasized the need for similar initiatives from the community, and pro-research data sharing policies in the advent of generative AI to remove barriers to accessibility without compromising data privacy and intellectual property. Beyond its data and technical contributions, UNIVERSALCEFR also carries broader sociolinguistic significance. UNIVERSALCEFR addresses the growing linguistic inequality in modern AI development by focusing on underrepresented languages alongside English. We hope this initiative can lead to more responsible AI development that actively resists the growing linguistic centralization around English in global AI research—a modern Matthew effect (Merton,1988)—where well-resourced languages receive disproportionate technological attention while smaller languages (like Czech or Welsh) are left behind (Masciolini et al.,2025). The UNIVERSALCEFR is a strong step towards mitigating the Matthew effect in language proficiency assessment research. Limitations We discuss several limitations of our work on UNIVERSALCEFR and how researchers can consider these directions to develop the resource further. Natural Data Disparity in Experiments. From the statistics presented in Tables 3and 7for UNIVERSALCEFR, it is expected that not all languages have the exact same distribution of data across dimensions, including formats (sentence-, paragraph- , document-, and dialogue-level) and category (reference and learner texts). Hence, our main experi9723 LANG SENT PARA DOC DIAG EN 12,826 409,362 1,837 0 ES 0 713 31,355 0 DE 26,244 1,033 5,673 0 NL 0 0 3,596 0 CS 0 441 0 0 IT 0 813 0 0 FR 1,669 0 344 0 ET 0 420 1,277 0 PT 0 1,423 0 0 AR 1,945 215 0 0 HI 1,491 0 0 0 RU 1,758 0 0 0 CY 1,107 109 41 115 Total 47,040 414,529 115 44,123 Table 7: Data statistics of UNIVERSALCEFR-FULL in terms of levels (sentence, paragraph, document, dialogue) across the 13 target languages. A Full Data Statistics Tables 7,9,11and 13 report the quantity of CEFRlabeled texts across granularity levels per language, and Tables 3,8,10 and 12 reflect their counterparts in terms of CEFR level coverage. In forming the TEST split, we randomly sampled CEFRlabeled text instances per language per granularity level, while setting a cap of 200. This allows us to have a sizeable representation of UNIVERSALCEFR while maintaining efficiency for inference with LLMs. In total, we have 4,465 CEFR-labeled instances for UNIVERSALCEFR-TEST, which is comparable to the general sizes of benchmark test sets from previous works related to language proficiency (Naous et al.,2024;Zhang et al.,2024; Imperial and Tayyar Madabushi,2024). For the TRAIN and DEV sets for fine-tuning and featurebased classification, we split the FULL subset (minus the TEST set) into a 90%-10% partition, respectively. B Coverage of Large Language Models In Table 14, we map each model’s language coverage or language support based on its respective release papers and publications. Language support means what specific languages have been added and in substantial quantities in a model’s training data (e.g., multilingual Wikipedia data dumps for pretraining XLM-R (Conneau et al.,2020)). LANG A1 A2 B1 B2 C1 C2 EN 173,005 119,335 59,634 20,746 7,122 675 ES 4577 4989 4,051 3,007 1,707 0 DE 273 13,208 12,996 346 108 308 NL 18 93 323 277 84 33 CS 1 92 77 38 2 0 IT 17 261 267 1 0 0 FR 106 302 404 335 210 98 ET 0 266 406 293 215 0 PT 204 62 270 59 80 0 AR 62 207 407 445 285 153 HI 203 219 223 203 182 145 RU 327 234 331 256 192 69 CY 463 332 0 0 0 0 Total 179,256 139,600 79,389 26,006 10,187 1,481 Table 8: Data statistics of UNIVERSALCEFR-TRAIN in terms of recognized CEFR levels (A1, A2, B1, B2, C1, C2) across the 13 target languages. C Language-Specific Analysis We provide in-depth analysis of model performances from the experiments in Section 5across multiple dimensions of UNIVERSALCEFR on results for selected languages that we are qualified to interpret. English. Analysis of model performance shows that using fine-tuned models and linguistic feature-based classification (62%-75%) obtains the best performance compared to prompting with instruction-tuned LLMs (19%-28%). However, these models tend to provide distinct patterns of specific CEFR labels. For the prompting setup, Gemma1, Gemma3, and EuroLLM models tend to give labels within the A1 and B1 range, while fine-tuned and feature-based models tend to lean towards the B1 and B2 range. For the pre-trained and instruction-tuned models, this finding may be tied to A1 and B2 being the most common CEFR level band of most general-purpose texts found online, where the sources of the data from which these models are trained. For feature-based models, we note the potential effect of training and test data having higher instance counts for these level bands than A1, C1, and C2. Regarding model scale, upgraded versions from similar model families perform better than their previous versions, echoing previous findings in literature (Imperial and Tayyar Madabushi,2024). This is particularly evident in Gemma3 being 12B in size and trained with massively multilingual data in 9730 LANG SENT PARA DOC DIAG EN 12,826 409,362 1,837 0 ES 0 713 31,355 0 DE 26,244 1,033 5,673 0 NL 0 0 3,596 0 CS 0 441 0 0 IT 0 813 0 0 FR 1,669 0 344 0 ET 0 420 1,277 0 PT 0 1,423 0 0 AR 1,945 215 0 0 HI 1,491 0 0 0 RU 1,758 0 0 0 CY 1,107 109 41 115 Total 47,040 414,529 115 44,123 Table 9: Data statistics of UNIVERSALCEFR-TRAIN in terms of levels (sentence, paragraph, document, dialogue) across the 13 target languages. 140+ languages and obtaining 28% in weighted F1 compared to Gemma1, which is 7B in size and English-centric, obtaining 21.8%. We note a potential default effect in using these models where additional specific CEFR descriptor information is not needed if the texts being evaluated are in English, due to the majority of data in the context of CEFR that is reflected in the training data being English. Spanish. Fine-tuned models outperform other setups, with feature-based approaches, especially Random Forest, achieving reasonable comparative performance. Moreover, multilingual models provide noticeable performance gains when compared to the English-only model. As per prompting strategy, for smaller multilingual models the language-specific prompt seems to play a role in improving the performance as it also does for the Gemma1 English-only model, however, the Gemma3 with 12B parameter is not affected by this, and it has been able to produce the best results of the LLMs (plus more sophisticated prompting strategies). As for the granularity of the input, models perform noticeably better at the document level than at the paragraph level, indicating that longer contexts are easier to classify than short ones. Finally, it is worth reporting a noticeable error of Gemma1: the prediction of C2 grade level, which does not exist in the Spanish dataset. Hindi. Both the Gemma models perform poorly LANG A1 A2 B1 B2 C1 C2 EN 19,449 13,151 6,643 2,384 797 85 ES 1535 1226 904 471 285 0 DE 32 2,494 2,392 60 13 41 NL 6 70 235 230 99 32 CS 0 14 9 6 0 0 IT 3 33 23 1 0 0 FR 13 30 39 43 20 12 ET 0 19 52 21 25 0 PT 61 213 50 144 19 61 AR 7 26 56 53 35 15 HI 22 30 20 16 12 13 RU 34 23 25 34 21 9 CY 67 44 0 0 0 0 Total 21,229 17,373 10,448 3,463 1,326 268 Table 10: Data statistics of UNIVERSALCEFR-DEV in terms of recognized CEFR levels (A1, A2, B1, B2, C1, C2) across the 13 target languages. compared to the fine-tuned XLM-R and the Random Forest variants and tend to classify most Hindi test items as A1 or A2. For example, Gemma1 puts 57% of Hindi test samples as A1, whereas there are only 19% of the test samples labeled as A1 in the gold standard labels. This is in line with the general trend noticed in Section 6.2, as the Hindi subset is entirely sentence-level. The distribution is closer to the Gold distribution for the fine-tuned and feature-engineered models. XLM-R fine-tuned models give the best performance amongst all models for Hindi, both in terms of exact category prediction and in terms of the degree of error (i.e., being within 1 level above or below the correct level). Finally, we looked at the correlation between a simple approximation of text length (calculated as the number of space-separated tokens), a commonly used variable in such automated language assessment approaches in NLP research, and the CEFR gold labels, as well as model-predicted labels, after converting them to a numeric scale. There was a high correlation between text length and the gold labels (0.7), which was also seen with the XLM-R model (0.74) and the Random Forest models (0.77). However, the Gemma models only had correlations of 0.44 and 0.54, respectively, with text length. However, considering that the Hindi subset only has sentence-level annotations without a larger context, it may be challenging to achieve further consistency with the gold standard labels, given the size of the annotated dataset. Future 9731 LANG SENT PARA DOC DIAG EN 1,274 40,980 0 255 ES 0 51 0 4,370 DE 4,168 79 0 785 NL 0 0 0 672 CS 0 29 0 0 IT 0 60 0 0 FR 146 0 0 11 ET 0 19 0 98 PT 0 548 0 0 AR 188 4 0 0 HI 113 0 0 0 RU 146 0 0 0 CY 111 0 0 0 Total 6,146 41,770 0 6,191 Table 11: Data statistics of UNIVERSALCEFR-DEV in terms of levels (sentence, paragraph, document, dialogue) across the 13 target languages. research should expand the available CEFR-graded resources both in terms of quantity as well as granularity for the language. Russian. The Russian results follow the broad patterns reported in the paper, but their rich inflectional morphology and their comparatively limited training data amplify several effects. Gemma1 (34.8%) greatly over-predicts texts as beginner-level (only 5% of texts had predictions above B1), confirming the overall trend that small, English-centric LLMs struggle most with morphologically rich languages. Gemma3 (37.4%) partially corrects this, but still massively underpredicts B2 and C2. XLM-R (49.6%) mirrors the gold distribution most faithfully, possibly because its multilingual vocabulary gives it better coverage of Russian inflectional morphology, a pattern also seen for other highly inflected languages such as Czech. The two Random Forest models (47.2% and 47.8%) under-predict A2 and C2 but otherwise match the gold shape, showing that handcrafted lexical and morpho-syntactic features capture useful Russian-specific signals even with limited data. Subword-level multilingual models (XLM-R) or explicit morpho-syntactic features (RF) are best suited to capture the meanings and relations between Russian words. Text length appears to be a false friend; although it does correlate highly with readability (r=0.65), it also appears to be the source of many errors; top-performing model outputs had text length correlations as high LANG A1 A2 B1 B2 C1 C2 EN 107 114 132 129 83 35 ES 49 58 140 108 45 0 DE 14 264 238 67 9 8 NL 4 21 69 77 22 7 CS 0 82 79 37 2 0 IT 9 87 104 0 0 0 FR 32 57 132 100 63 16 ET 0 110 130 93 67 0 PT 49 50 47 30 13 11 AR 12 26 162 145 40 15 HI 38 34 42 42 28 16 RU 41 36 52 35 24 12 CY 233 232 0 0 0 0 Total 588 1,171 1,327 863 396 120 Table 12: Data statistics of UNIVERSALCEFR-TEST in terms of recognized CEFR levels (A1, A2, B1, B2, C1, C2) across the 13 target languages. as 0.73. Since this experiment with Russian is limited to sentence-level readability, comparison with previous research on Russian readability assessment is not straightforward. However, the weighted F1 (49.6%) of the best-performing model (XLM-R) is below state-of-the-art results for longer texts, including 67% (Reynolds,2016), 74% (Solnyshkina et al.,2018), and 78% (Blinova and Tarasov,2022). Most likely, this difference is partly due to the absence of Russian-specific morphosyntactic features that have been highly informative in previous studies’ models. Portuguese. Comparing the different setups, we can see that the results for Portuguese follow the global tendency, with fine-tuned models achieving the highest performance, followed by feature-based models, and with prompting taking the last place. Although this study only covers paragraph-level learner data for Portuguese, similar patterns were observed on reference data (Ribeiro et al.,2024b). However, comparing the results with those of other languages and, particularly, those with paragraph-level learner data, we can see that Portuguese is the language with the lowest performance ( ≈ 33.5%). Several factors may contribute to this outcome. For instance, Portuguese is one of the languages with the least available training data, and the distribution of proficiency labels is right-skewed (especially in COPLE2). Furthermore, the data consists of texts written by learners from a wide range of 9732 LANG SENT PARA DOC DIAG EN 200 200 200 0 ES 0 200 200 0 DE 200 200 200 0 NL 0 0 200 0 CS 0 200 0 0 IT 0 200 0 0 FR 200 0 200 0 ET 0 200 200 0 PT 0 200 0 0 AR 200 200 0 0 HI 200 0 0 0 RU 200 0 0 0 CY 200 109 41 115 Total 1,400 1,709 1,241 115 Table 13: Data statistics of UNIVERSALCEFR-TEST in terms of levels (sentence, paragraph, document, dialogue) across the 13 target languages. L1 backgrounds with generally low proficiency. This makes it more difficult for models to identify consistent patterns due to strong L1 interference and low coverage. Overall, both fine-tuned and feature-based models seem to be unable to distinguish between sublevels, with most examples of both A levels being predicted as A1, and the remainder (mostly examples of the B levels) as B1. On the positive side, contrary to what was observed for other languages, the models do not seem to be influenced by text length, with the predictions of XML-R having a correlation of just 0.39 with that feature. The prompting approaches lead to a bias towards the prediction of levels A2 and B1, with the top performer among these approaches (Gemma3 with EN-WRITE prompt) predicting A2 for 28% of the examples and B1 for 62%. Notably, when using the more descriptive prompts, the Gemma 1 model outperformed EuroLLM, in spite of having fewer parameters and not being specifically trained on Portuguese data. French. The French corpus and our analysis are divided into sentence-level and document-level data. The sentence-level set contains 1,668 sentences ranging from A1 to C2, while the document-level set includes 344 documents from A1 to C1, with an intense concentration at the B levels (75% of the data falls within B1 and B2). In line with the other languages, XLM-R is the most consistent model and achieves the best global performance in every setting. Random Forest (RF) with all features fluctuates more in overall performance, dropping notably in the document-level task, but retains some consistency in terms of which proficiency levels it performs best or worst on. RF with top features performs inconsistently overall but achieves the best results on the document-level task. However, it shows instability in class-level performance, with changes in which levels are most accurately predicted. Among the prompt-based models, Gemma3 is more stable than Gemma1, but both remain below the performance of XLM-R and RF, showing a weaker performance in the LLMs (Gemma1 and Gemma3). Gemma1, in particular, is the least consistent model, with highly variable class-level performance and occasional zero F1 scores for some levels in specific setups. The Gemma1 results are likely due to the lack of French documents during the training of this model. Across all models, prediction is generally more reliable for intermediate levels (A2–B2), while C-level predictions remain the most challenging. Fine-tuning has the clear advantage: the fine-tuned XLM-R achieves the highest accuracy across all evaluation set-ups, making it the most reliable in correctly predicting gold labels. It consistently outperforms all other models, both at the sentence and document levels. This is consistent with previous experiments on French (Yancey et al., 2021;Ngo and Parmentier,2023;Wilkens et al., 2024), although our performance is slightly lower than in those studies. Prompting is the least effective: both Gemma1 and Gemma3, used in a prompt-based setting, show the lowest prediction accuracy, often failing to identify the correct labels, especially at the extremes of the proficiency scale (A1, C1, and C2 levels). Traditional supervised classifiers (Random Forest) perform moderately well, consistently outperforming the prompt-based models but still lagging behind the fine-tuned model. The feature-based models had a particularly poor performance on C1 and C2 levels. This is likely due to a lack of specialized features for those proficiency levels. Moreover, their performance varies by set-up, with some gains at the document level but noticeable drops elsewhere. Nevertheless, the two RF flavours had similar results. In summary, fine-tuning yields the best predictions, followed by traditional supervised learning, while prompting underperforms in this task. 9733 Model EN ES DE NL CS IT FR ET PT AR HI RU CY Tally GEMMA1✓1/1 GEMMA3✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ 13/140 EUROLLM ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ 12/35 MODERNBERT ✓1/1 EUROBERT ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ 10/15 XLM-R ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ 13/100 Table 14: Mapping of language coverage of training data used for the six large, pretrained language models in the model evaluation paradigm in Section 5. Models in teal are English-centric (trained primarily with English data), and models in purple are multilingual (trained with massive multilingual data). We referred to each model’s corresponding release papers and publications for information on their supported languages. Note that the documentation of GEMMA3 indicates it has been trained with 140+ languages. Thus, we loosely consider it to cover all 13 languages in UNIVERSALCEFR. The tally column indicates { lang_covered / lang_seen }. For example, EUROBERT covers 10 of the languages in the current UNIVERSALCEFR the 15 languages it supports. German. For German, the fine-tuned models (>70%) have been shown to outperform all other approaches, such as feature-based ( ≈ 50%-65%) and prompting ( ≈ 38%-46%), despite the presence of unbalanced CEFR levels in both the training and test data. The findings derived from the English-only and multilingual models, including fine-tuning and prompting methodologies, exhibit no notable difference. This may be due to the similarities between English and German, both of which are West Germanic languages. Alternatively, the great transferability of the fine-tuned Englishonly model may also be due to the large amount of German training data available (27,000 training samples). The feature-based models performed second best and were still able to compete with the fine-tuned models to some extent. This is surprising, given that a previous analysis showed that the features only exhibited low correlations with CEFR levels (see Section E.3). Proficiency assessment for German appears to require certain idiosyncratic features. For example, the feature covering the maximum distance between words in a dependency tree showed a high feature importance only for German, reflecting the language’s free word order and long-distance dependencies. For the prompting setup, the multilingual Gemma3 model performed, achieving good results for lower CEFR levels, but underpredicting higher levels. By contrast, Gemma1 significantly overpredicts level A1 (250 against 14 from the gold labels), resulting in poorer performance on average and across the other levels. One deceptive indicator might be the length of the texts to be classified, as reflected by the strong correlation between text length and Gemma1’s predictions ( r =0.61). When comparing the prompting setups with regard to language-specific task descriptions, no clear trend emerges across all three LLMs, mirroring the difficulty of prompt engineering for a complex task such as multi-lingual proficiency classification. Arabic. Across the 400 Arabic test items, Gemma1 tends to over-predict lower CEFR levels, assigning 31 items to A1 while only 12 are from the true labels, and 90 to A2 against 26. There is also a tendency to under-predict C1, with 18 predictions against 40 from the true labels, resulting in the highest average grade deviation of 1.0. In contrast, XLM-R and both Random Forest variants distributed their predictions more evenly overall, with XLM-R achieving the smallest average grade deviation of 0.75. In terms of granularity, the Arabic subset is split into sentence-level, reference data, and paragraph-level learner data. For the sentencelevel reference texts, XLM-R ( ≈ 55%) and Random Forest models from the two linguistic feature setups ( ≈ 49.3%-51.2%) outperform both Gemma1 and Gemma3 models through prompting ( ≈ 16.5%- 32%). However, with paragraph-level learner texts, Gemma3 leads the evaluation ( ≈ 41%). At the same time, XLM-R and the Random Forest models fall behind ( ≈ 32%), possibly due to the Arabic data used in the training split, which are entirely sentence-level. In contrast, the Gemma3 model has most likely seen diverse online Arabic data. D Standardized Dataset Fields We present the standardized JSON format used as a template when processing all qualified datasets in UNIVERSALCEFR. This structured format ensures flexibility and interoperability into other formats 9734 accepted and used by the AI community, including Huggingface and Croissant. Moreover, this format captures the dimensions that are essential to each instance of CEFR-labeled text, including format or granularity, category, license, and language. E Full Linguistic Feature Analysis E.1 All Linguistic Features Overall, we have extracted 100 diverse linguistic features which can be grouped into morphosyntactic (62), syntactic (18), length-based (11), lexical (4), readability (2), psycholinguistic (2), and discourse (1). The full list of features, including short descriptions, is available in Appendix E. We extracted a diverse set of 100 linguistic features based on sentence-based linguistic annotation with spacy (Montani et al., 2023) and stanza (Qi et al.,2020), including tokenization, part-of-speech tagging, and dependency parsing performed. Additionally, we use fasttext embeddings (Grave et al.,2018), pyphen for hyphenation (Berendsen and Kozea, 2025) and MEGA.HR crossling lexicon 9 for imageability and concreteness (Ljubeši´ c,2018). Most of the features have already been implemented in the text-simplification-evaluation (TSEval) package 10 (see Martin et al. (2018) for the original version and Stodden and Kallmeyer (2020) for the multilingual version). In Table 21, we provide an overview of all features including a short description, resources used, and correlation with the CEFR level. E.2 Top Linguistic Features To extract the top linguistic features (TOPFEATS), we selected those that are present in the top 10 ranked most important features for at least three languages. Using this criteria, we came up with a list of 23 linguistic features as reported in Table 16 which was then used in the experiment result in Table 4. E.3 Linguistic Correlation Analysis In the following, we describe some insights into linguistic diversity of the UniversalCEFR data by correlation analysis between the features and the CEFR levels. 9https://www.clarin.si/repository/xmlui/handl e/11356/1187 10https://github.com/facebookresearch/text-sim plification-evaluation Correlation Across All Languages. Considering the absolute Spearman correlation between the features and the CEFR level (selecting values with p < 0.05 and ρ > 0.3 on average across all languages), the strongest associations were found in length-based measures, such as characters per sentence and syllables per sentence. Several grammatical complexity features, including parse tree height and phrase length, showed moderate correlations. Readability indices (FKGL and Flesch Reading Ease) also displayed moderate correlations in the expected direction. Psycholinguistic features, such as concreteness and imageability, were negatively correlated with proficiency, indicating a shift toward more abstract language at higher levels. Finally, morphosyntactic features regarding voice, tense, and number showed moderate but consistent correlations, supporting their relevance in reflecting syntactic development. Correlation By CEFR Level. To assess the consistency of feature relevance across languages, we examined the number of features with significant correlations ( p < 0.05 ) with CEFR levels per language. The results revealed notable variations. Languages such as Czech (CS), Estonian (ET), and Italian (IT) showed a high number of relevant features, suggesting strong alignment between the selected linguistic features and CEFR progression in these languages. English (EN), Spanish (ES), French (FR), Hindi (HI), and Russian (RU) showed moderate coverage, with a reasonable number of features exceeding the 0.3 correlation threshold. In contrast, Arabic (AR), Dutch (NL), and Portuguese (PT) exhibited weak coverage, while Welsh (CY) and German (DE) had very few or no features with relevant correlations, indicating a limited match between the current feature set and CEFR levels for those languages. Furthermore, a few features are only relevant for a few languages, e.g., the translative case for only Estonian, negative verb polarity for only Czech, or genitive case for only Czech, Estonian, and Russian. This variability highlights the influence of language-specific properties on the effectiveness of general feature-based models for proficiency prediction. Point-Biserial Correlation. A point-biserial correlation analysis by CEFR level revealed that most features exhibit only weak correlations, suggesting limited discriminative power when isolating individual CEFR bands. Interestingly, the absolute correlation values tend to be strongest at the 9735 Field Description title The unique title of the text retrieved from its original corpus ( NA if there are no titles such as CEFR-assessed sentences or paragraphs). lang The source language of the text in ISO 638-1 format (e.g., en for English). source_name The source dataset name where the text is collected as indicated from their source dataset, paper, and/or documentation (e.g., cambridge-exams from Xia et al. (2016)). format The format of the text in terms of level of granularity as indicated from their source dataset, paper, and/or documentation. The recognized formats are the following: [ document-level , paragraph-level , discourse-level , sentence-level]. category The classification of the text in terms of who created the material. The recognized categories are reference for texts created by experts, teachers, and language learning professionals and learner for texts written by language learners and students. cefr_level The CEFR level associated with the text. The six recognized CEFR levels are the following: [ A1, A2, B1, B2, C1, C2 ]. A small fraction (<1%) of text in UNIVERSALCEFR contains unlabelled text, texts with plus signs (e.g., A1+), and texts with no level indicator (e.g., A, B). license The licensing information associated with the text ( Unknown if not stated). text The actual content of the text itself. Table 15: The structured JSON fields with descriptions and examples used as the standardized uniform format for building the UNIVERSALCEFR dataset. All instances validated from the collection of CEFR-labelled corpora conform to this format. 9736 CATEGORY FEATURE NAME Length doc_num_sents doc_num_tokens num_characters num_characters_per_sentence num_characters_per_word num_syllables_in_sentence num_syllables_per_sentence num_syllables_per_word num_words Lexical average_pos_in_freq_table lexical_complexity_score Morphosyntactic ratio_Tense_Past ratio_of_determiners ratio_of_numerals ratio_of_pronouns Psycholinguistic concreteness imagebility Readability sentence_fkgl sentence_fre Syntactic avg_distance_between_words average_length_VP parse_tree_height ratio_of_coordinating_clauses Table 16: List of linguistic features occurring in the top 10 of at least three languages. We use this list for the TOPFEATURES subset used in the experiment result in Table 4. A1 level, particularly for psycholinguistic features such as imageability ( ρ= 0.48 ) and concreteness ( ρ= 0.46 ), as well as punctuation-related measures. This suggests that certain surface-level and lexical-semantic features may be especially informative at the lowest proficiency level. A notable case is the feature of word length in characters, which shows a negative correlation at A1 ( ρ=−0.45 ), becomes neutral at A2, and shifts to a positive correlation at B1 and higher levels. This pattern may reflect increasing lexical complexity with proficiency. Similarly, features related to syntactic structure, such as the ratio of past tense verbs and phrase length, generally shift from weak negative to weak positive correlations as proficiency increases, indicating progressive syntactic development. Overall, the directionality of several features suggests dynamic usage patterns across CEFR bands, even if the correlation strengths remain modest. F Hyperparameter Values We detail the hyperparameter values used for finetuning pretrained (MODERNBERT, EUROBERT, HYPERPARAMETER VALUE Learning rate 3.6×10−5 Train batch size 2 Evaluation batch size 3 Random seed 42 Gradient accumulation steps 16 Total effective batch size 32 Optimizer adamw_torch_fused Betas (0.9, 0.999) Epsilon 10−8 Learning-rate scheduler linear Warm-up ratio 0.1 Table 17: Hyperparameter values used for fine-tuning pretrained language models. HYPERPARAMETER VALUE Sampling False Max New Tokens 10 Data Type torch.bfloat16 GPU 4 x NVIDIA RTX A5000 (24GB) Table 18: Hyperparameter values and GPU information used for prompting instruction-tuned models. LANG SENT PARA DOC OVERALL AR 55.7 32.6 - 43.1 CY 86.9 72.5 61.5 72.7 CS - 68.8 - 68.8 DE 65.4 71.1 83.4 73.2 EN 68.3 100.0 57.6 75.5 ES - 40.6 98.0 69.69 ET -93.6 84.0 88.9 FR 57.6 - 44.2 51.7 HI 52.9 - - 52.9 IT - 83.3 - 83.3 NL - 59.0 59.0 PT - 29.2 - 29.2 RU 49.6 - - 49.6 Table 19: Weighted F1 scores for the fine-tuned XLMR (top model across all setups) performance on the UNIVERSALCEFR-TEST, classified by the granularity levels of the data. and XLM-R) and instruction-tuned language models (GEMMA1, GEMMA3, and EUROLLM) in Tables 17 and 18, respectively. 9737 G Additional Context on Restrictions of GDPR-Protected Datasets The critical aspect of the GDPR is that it gives data subjects (e.g., L2 learners of CEFR) the right to withdraw their personal information from processing, which requires data processors to store both the signed consents and the ID mappings (i.e., mappings between the names of the real people and their IDs in a released corpora). As long as these documents exist and reidentification is theoretically possible, the data falls under the scope of the GDPR. Further complicating factors are national legislations and ethical regulations, such as archival laws, that treat any data produced at universities— including those used for language proficiency assessment such as essays, recorded dialogues, and written texts from personal experiences—as the property of the state (and hence making destruction of the ID mappings a non-trivial act) (European Parliament and Council,2016). Yet another upcoming challenge is the EU AI Act (European Parliament and Council,2024) that implies that AI models trained on personal data should inherit the same license as the data they have been trained on, meaning that the models will be under the scope of the GDPR. We hypothesize that the non-restricted datasets included in UNIVERSALCEFR either do not contain personal information or were collected before the GDPR, since they are already openly accessible to the public. We further hypothesize that the datasets currently under the GDPR will eventually have their ID mappings destroyed and will no longer be subject to the GDPR. This may mean that the learner corpora that can be added to UNIVERSALCEFR will grow with time. H Full Dataset Directory of UniversalCEFR We provide the complete information of qualified corpora included in the current UNIVERSALCEFR collection to form a directory of datasets. Aside from eight per-instance information included in the standardized JSON format in Table 15, we also report five per-corpus information as listed below: • Annotation method used (manual, computerassisted, or NA). • Total number of expert annotators. • Distinct L1 learners per language for learner corpora. • Inter-annotator agreement (IAA) metric and score. • Reference to published paper or repository. I Prompt Templates We provide the complete copies of the prompt templates used in prompting experiments with instruction-tuned LLMs as described in Section 5. The prompt templates are categorized by color based on the setup: BASE,EN-READ,LANGREAD,EN-WRITE,LANG-WRITE. J Welsh Data Collection One of the contributions of UNIVERSALCEFR is the release of the first-ever open dataset for the Welsh language (CY) with gold-standard CEFR labels for A1 and A2. To obtain this data, we corresponded with data maintainers from Learn Welsh ( https://learnwelsh.cymru/ ), which is a compilation of expert-created books (reference texts) and acquired PDF versions. This resource can be shared in any format for non-commercial research, which fits the goal of UNIVERSALCEFR. We then manually extracted qualified texts according to the four levels of granularity: sentence, paragraph, dialogue, and document. The distribution of CEFR levels and text granularity for this new Welsh dataset can be found in Table 3and 7, respectively. 9738 Category Feature Short Description Resource Corr. ρ (avg.) Corr. ρ (SD) Discourse ratio_referential Ratio of referential tokens to all tokens based on dependency tree relations SpaCy, Stanza, TSEval 0.1278 0.09 Length doc_num_sents Number of sentences per document / text SpaCy, Stanza, 0.1819 0.18 doc_num_tokens Number of tokens per document / text SpaCy, Stanza, 0.5041 0.22 num_characters Number of characters per document / text SpaCy, Stanza, TSEval 0.5224 0.19 num_characters_per_sentence Number of characters per sentence SpaCy, Stanza, TSEval 0.5301 0.17 num_characters_per_word Number of characters per word SpaCy, Stanza, TSEval 0.3895 0.16 num_sentences Number of sentences per document / text SpaCy, Stanza, TSEval -0.0324 0.11 num_syllables_in_sentence Number of syllables in document / text SpaCy, Stanza, pyphen, TSEval 0.525 0.19 num_syllables_per_sentence Number of syllables per sentence SpaCy, Stanza, pyphen, TSEval 0.4634 0.23 num_syllables_per_word Number of syllables per word SpaCy, Stanza, pyphen, TSEval 0.3924 0.16 num_words Number of tokens per document / text SpaCy, Stanza, TSEval 0.479 0.18 num_words_per_sentence Number of tokens per sentence SpaCy, Stanza, TSEval 0.4863 0.16 average_pos_in_freq_table Average frequency rank of tokens in FastText embeddings SpaCy, Stanza, TSEval -0.0907 0.22 lexical_complexity_score Lexical complexity based on ranks of FastText embeddings SpaCy, Stanza, FastText, TSEval 0.0649 0.13 type_token_ratio Type-token-Ratio SpaCy, Stanza, TSEval -0.3364 0.16 Lexical max_pos_in_freq_table Maximum frequency rank of tokens in FastText embeddings SpaCy, Stanza, TSEval 0.2014 0.07 Psycholinguistic concreteness Concreteness of words based on MEGAHR and FastText-Embeddings MEGA.HR crossling -0.4254 0.24 imagebility imagebility of words based on MEGAHR and FastText-Embeddings MEGA.HR crossling -0.3962 0.24 sentence_fkgl Flesch-Kincaid-Grading-Level, designed for English SpaCy, Stanza, TSEval 0.4738 0.19 Readability sentence_fre Flesch-Reading Ease, designed for English SpaCy, Stanza, TSEval -0.3976 0.19 Syntactic avg_distance_betweeen_verb_particle Average distance between verb and particle based on dependency tree SpaCy, Stanza, 0.0385 avg_distance_betweeen_words Average distance between words based on dependency tree SpaCy, Stanza, 0.3934 0.17 max_distance_betweeen_verb_particles Maximum distance between verb and particle based on dependency tree SpaCy, Stanza, 0.0385 max_distance_betweeen_words Maximum distance between words based on dependency tree SpaCy, Stanza, 0.2109 0.14 check_if_head_is_noun Whether the head of the dependency tree is a noun SpaCy, Stanza, TSEval -0.1147 0.13 check_if_head_is_verb Whether the head of the dependency tree is a verb SpaCy, Stanza, TSEval 0.0954 0.13 check_if_one_child_of_root_is_subject Whether a child of a root is a subject (not a verb) SpaCy, Stanza, TSEval 0.0144 0.15 check_passive_voice Whether a sentence is in passive voice SpaCy, Stanza, TSEval average_length_NP Average length of noun phrase in tokens SpaCy, Stanza, TSEval 0.3485 0.13 average_length_VP Average length of verb phrase in tokens SpaCy, Stanza, TSEval 0.4324 0.16 avg_length_PP Average length of prepositional phrase in tokens SpaCy, Stanza, TSEval 0.1406 0.07 parse_tree_height Depth or height of the dependency tree SpaCy, Stanza, TSEval 0.4657 0.15 ratio_clauses Ratio of tokens associated to a clause to all tokens based on dependency tree relations SpaCy, Stanza, TSEval 0.1918 0.12 ratio_of_coordinating_clauses Ratio of tokens associated to a coordinating clause to all tokens based on dependency tree relations SpaCy, Stanza, TSEval 0.2503 0.16 ratio_of_subordinate_clauses Ratio of tokens associated to a subordinating clause to all tokens based on dependency tree relations SpaCy, Stanza, TSEval 0.2361 0.14 ratio_prepositional_phrases Ratio of tokens associated to a prepositional phrase to all tokens based on dependency tree relations SpaCy, Stanza, TSEval 0.1797 0.14 ratio_relative_phrases Ratio of tokens associated to a relative clause to all tokens based on dependency tree relations SpaCy, Stanza, TSEval 0.3262 0.15 is_non_projective Whether a dependency tree is non projective SpaCy, Stanza, TSEval 0.0883 0.13 ratio_Abbr_Yes Ratio of nouns which are an abbreviation to all nouns SpaCy, Stanza, UniversalDependencies -0.0685 0.07 ratio_Case_Abe Ratio of nouns in abessive case to all nouns SpaCy, Stanza, UniversalDependencies 0.2394 ratio_Case_Acc Ratio of nouns in accusative case to all nouns SpaCy, Stanza, UniversalDependencies 0.0658 0.29 ratio_Case_Ben Ratio of nouns in benefactive case to all nouns SpaCy, Stanza, UniversalDependencies ratio_Case_Cau Ratio of nouns in causative case to all nouns SpaCy, Stanza, UniversalDependencies ratio_Case_Cmp Ratio of nouns in comparative case to all nouns SpaCy, Stanza, UniversalDependencies ratio_Case_Cns Ratio of nouns in considerative case to all nouns SpaCy, Stanza, UniversalDependencies ratio_Case_Com Ratio of nouns in comitative case to all nouns SpaCy, Stanza, UniversalDependencies 0.1293 ratio_Case_Dat Ratio of nouns in dative case to all nouns SpaCy, Stanza, UniversalDependencies 0.1822 0.09 ratio_Case_Dis Ratio of nouns in distributive case to all nouns SpaCy, Stanza, UniversalDependencies ratio_Case_Equ Ratio of nouns in equative case to all nouns SpaCy, Stanza, UniversalDependencies Morphosyntactic ratio_Case_Erg Ratio of nouns in ergative case to all nouns SpaCy, Stanza, UniversalDependencies Table 20: Overview of all 100 features, including correlation coefficient with CEFR level across all languages. 9739 CEFR specifications for reading comprehension in Czech (CS) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Czech text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 – Studenti této úrovnˇ e dokážou porozumˇ et velmi krátkým jednoduchým text˚um po jedné frázi, pochytají známá jména, slova a základní fráze a pˇ reˇ ctou si je podle potˇ reby. A2 – Studenti této úrovnˇ e dokážou porozumˇ et krátkým jednoduchým text˚um obsahujícím nejfrekventovanˇ ejší slovní zásobu, vˇ cetnˇ eˇ cásti sdílené mezinárodní slovní zásoby. B1 – Studenti této úrovnˇ e dokážou ˇ císt pˇ rímoˇ caré vˇ ecné texty na témata související s jejich oblastí zájmu s uspokojivou úrovní porozumˇ ení. B2 – Studenti této úrovnˇ e dokážou ˇ císt s velkou mírou nezávislosti, pˇ rizp˚usobují styl a rychlost ˇ ctení r˚uzným text˚um a úˇ cel˚um a selektivnˇ e používají vhodné referenˇ cní zdroje. Má širokou slovní zásobu aktivního ˇ ctení, ale m˚uže mít potíže s nízkofrekvenˇ cními idiomy. C1 – Studenti této úrovnˇ e dokážou podrobnˇ e porozumˇ et dlouhým a složitým text˚um, at’ už se týkají nebo netýkají jejich vlastní oblasti specializace, za pˇ redpokladu, že dokážou znovu pˇ reˇ císt obtížné ˇ cásti. Mohou také porozumˇ et široké škále text˚u, vˇ cetnˇ e literárních text˚u, ˇ clánk˚u v novinách nebo ˇ casopisech a specializovaných akademických nebo odborných publikací, za pˇ redpokladu, že mají pˇ ríležitosti k opakovanému ˇ ctení a mají pˇ rístup k referenˇ cním nástroj˚um. C2 – Studenti této úrovnˇ e mohou porozumˇ et prakticky všem typ˚um text˚u vˇ cetnˇ e abstraktních, strukturálnˇ e složitých nebo vysoce hovorových literárních a neliterárních spis˚u. Dokážou také porozumˇ et široké škále dlouhých a složitých text˚u, ocenit jemné rozdíly ve stylu a implicitní i explicitní význam. Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 9746 CEFR specifications for reading comprehension in Italian (IT) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Italian text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 - Gli studenti di questo livello riescono a comprendere testi molto brevi e semplici, una frase alla volta, cogliendo nomi familiari, parole e frasi di base e rileggendo quando necessario. A2 - Gli studenti di questo livello riescono a comprendere testi brevi e semplici contenenti il vocabolario più frequente, inclusa una parte di elementi di vocabolario internazionale condiviso. B1 - Gli studenti di questo livello riescono a leggere testi fattuali semplici su argomenti correlati al loro campo di interesse con un livello di comprensione soddisfacente. B2 - Gli studenti di questo livello riescono a leggere con un ampio grado di indipendenza, adattando stile e velocità di lettura a testi e scopi diversi e utilizzando fonti di riferimento appropriate in modo selettivo. Ha un ampio vocabolario di lettura attiva, ma può avere qualche difficoltà con idiomi a bassa frequenza. C1 - Gli studenti di questo livello riescono a comprendere in dettaglio testi lunghi e complessi, indipendentemente dal fatto che siano correlati o meno alla propria area di specializzazione, a condizione che riescano a rileggere sezioni difficili. Possono anche comprendere un’ampia varietà di testi, tra cui scritti letterari, articoli di giornali o riviste e pubblicazioni accademiche o professionali specializzate, a condizione che vi siano opportunità di rilettura e abbiano accesso a strumenti di riferimento. C2 - Gli studenti di questo livello possono comprendere praticamente tutti i tipi di testi, tra cui scritti letterari e non letterari astratti, strutturalmente complessi o altamente colloquiali. Possono anche comprendere un’ampia gamma di testi lunghi e complessi, apprezzando sottili distinzioni di stile e significato implicito ed esplicito. Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 9747 CEFR specifications for reading comprehension in French (FR) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given French text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 - Les apprenants de ce niveau peuvent comprendre des textes très courts et simples, phrase par phrase, en reprenant des noms, des mots et des expressions de base familiers et en les relisant si nécessaire. A2 - Les apprenants de ce niveau peuvent comprendre des textes courts et simples contenant le vocabulaire le plus courant, y compris une partie du vocabulaire international commun. B1 - Les apprenants de ce niveau peuvent lire des textes factuels simples sur des sujets liés à leur domaine d’intérêt avec un niveau de compréhension satisfaisant. B2 - Les apprenants de ce niveau peuvent lire avec une grande autonomie, en adaptant leur style et leur vitesse de lecture à différents textes et objectifs, et en utilisant sélectivement des sources de référence appropriées. Possède un vocabulaire de lecture actif et étendu, mais peut éprouver des difficultés avec les expressions idiomatiques peu fréquentes. C1 - Les apprenants de ce niveau peuvent comprendre en détail des textes longs et complexes, qu’ils relèvent ou non de leur domaine de spécialité, à condition de pouvoir relire les passages difficiles. Ils peuvent également comprendre une grande variété de textes, notamment des écrits littéraires, des articles de journaux ou de magazines, ainsi que des publications universitaires ou professionnelles spécialisées, à condition de disposer d’opportunités de relecture et d’outils de référence. C2 - Les apprenants de ce niveau peuvent comprendre pratiquement tous les types de textes, y compris les écrits littéraires et non littéraires abstraits, structurellement complexes ou très familiers. Ils peuvent également comprendre un large éventail de textes longs et complexes, en appréciant les subtilités stylistiques et le sens implicite et explicite. Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 9748 CEFR specifications for reading comprehension in Estonian (ET) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Estonian text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 – selle taseme õppijad saavad aru väga lühikestest lihtsatest tekstidest ühe fraasi kaupa, korjavad üles tuttavad nimed, sõnad ja põhifraasid ning loevad vajaduse korral uuesti läbi. A2 – selle taseme õppijad saavad aru lühikestest lihtsatest tekstidest, mis sisaldavad kõige sagedamini kasutatavat sõnavara, sealhulgas osa jagatud rahvusvahelistest sõnavaraüksustest. B1 – selle taseme õppijad oskavad rahuldaval mõistustasemel lugeda otsekoheseid faktitekste nende huvivaldkonnaga seotud teemadel. B2 – selle taseme õppijad oskavad lugeda suurel määral iseseisvalt, kohandades lugemisstiili ja -kiirust erinevate tekstide ja eesmärkidega ning kasutades valikuliselt sobivaid viiteallikaid. Tal on lai aktiivse lugemise sõnavara, kuid tal võib esineda raskusi madala sagedusega idioomidega. C1 – selle taseme õppijad saavad üksikasjalikult aru pikkadest ja keerukatest tekstidest, olenemata sellest, kas need on seotud nende enda erialaga või mitte, eeldusel, et nad suudavad raskeid lõike uuesti lugeda. Nad saavad aru ka paljudest erinevatest tekstidest, sealhulgas kirjanduslikest kirjutistest, ajalehtede või ajakirjade artiklitest ning erialastest akadeemilistest või erialastest väljaannetest, eeldusel, et neil on võimalus uuesti lugeda ja neil on juurdepääs viitevahenditele. C2 – selle taseme õppijad saavad aru peaaegu igat tüüpi tekstidest, sealhulgas abstraktsetest, struktuurselt keerukatest või väga kõnekeelsetest kirjanduslikest ja mittekirjanduslikest kirjutistest. Samuti saavad nad aru paljudest pikkadest ja keerulistest tekstidest, mõistes peent stiilieritlust ning kaudset ja selgesõnalist tähendust. Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 9749 CEFR specifications for reading comprehension in Portuguese (PT) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Portuguese text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 - Os alunos deste nível podem entender textos muito curtos e simples, uma única frase de cada vez, pegando nomes, palavras e frases básicas familiares e relendo conforme necessário. A2 - Os alunos deste nível podem entender textos curtos e simples contendo o vocabulário de maior frequência, incluindo uma proporção de itens de vocabulário internacional compartilhados. B1 - Os alunos deste nível podem ler textos factuais diretos sobre assuntos relacionados ao seu campo de interesse com um nível satisfatório de compreensão. B2 - Os alunos deste nível podem ler com um alto grau de independência, adaptando o estilo e a velocidade de leitura a diferentes textos e propósitos, e usando fontes de referência apropriadas seletivamente. Tem um amplo vocabulário de leitura ativa, mas pode ter alguma dificuldade com expressões idiomáticas de baixa frequência. C1 - Os alunos deste nível podem entender em detalhes textos longos e complexos, estejam eles relacionados ou não à sua própria área de especialidade, desde que possam reler seções difíceis. Eles também podem entender uma grande variedade de textos, incluindo escritos literários, artigos de jornais ou revistas e publicações acadêmicas ou profissionais especializadas, desde que haja oportunidades de releitura e tenham acesso a ferramentas de referência. C2 - Alunos deste nível podem entender virtualmente todos os tipos de textos, incluindo escritos abstratos, estruturalmente complexos ou altamente coloquiais, literários e não literários. Eles também podem entender uma grande variedade de textos longos e complexos, apreciando sutis distinções de estilo e significado implícito e explícito. Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 9750 CEFR specifications for reading comprehension in Arabic (ar) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Arabic text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: 󰁀 󰆳󰆎 ﺓﺍﻭ 󰃎󰃆󰄨󰄚 ،ﻭ ﺍً ﺓ ﺹ  ﻯﺍ ﺍ 󰆳󰆎 󰋜󰊈  – A1 .ﺍ  ﺓﺀﺍﺍ ﺓﺩ󰌣ﻭ ،ﺍ ﺍ ﺕﺍﺭﺍﻭ ﺕ󰀖󰀞ﺍﻭ ﺀ󰄺󰄫ﺍ ﻁﺍﻭ ،ﺓ ﺕﺍﺩﺍ ﺃ  ﻱ ﻭ ﺓ ﺹ  ﻯﺍ ﺍ 󰆳󰆎 󰋜󰊈  – A2 .󰂯󰂋ﺍ ﻭ󰃕󰃀ﺍ ﺕﺍﺩﺍ  ﺀ 󰊜󰊈ﺫ 󰆳󰆎  ،ً ﻝ  ﺍ ﻝ ﺓ ﺍﻭ ﺹ ﺓﺀﺍ ﻯﺍ ﺍ 󰆳󰆎 󰋜󰊈  – B1 .ﺍ  ٍﺽ ﻯ 󰉼󰓜ﺍ ﺏﺃ ﻭ ،ﺍ   ﺭ ﺓﺀﺍﺍ ﻯﺍ ﺍ 󰆳󰆎 󰋜󰊈  – B2 󰁀󰀞 ﺍ ﺍﺍ ﺭﺩ ﻡﺍﺍﻭ ،ﺽﺍﺍﻭ ﺹﺍ ﻑ ًﻭ ﺓﺀﺍﺍ ﻭ ﺍ  ﺍ  ﻥﺍ  ـ ،ﺍﻭ  ﺓﺀﺍ ﺕﺍﺩ 󰉼󰈸󰃕󰃀 .󰆳󰅗ﺍ .ﻡﺍﺍ ﺓﺭﺩ ﺍ 󰁅 ﺀﺍ ، ﺓﻭ 󰃎󰃆 ﺹ  ﻯﺍ ﺍ 󰆳󰆎 󰋜󰊈  – C1 .ﺍ ﺍ ﺓﺀﺍ ﺓﺩﺇ  ﺭﺩ ﺍ ﻥﺃ ﻁ ، ﻡﺃ 󰉾 ﻝ  ﺍ ﺕﻭ ،ﺩﺍ ﻝ󰄨󰄇ﺍ 󰊜󰊈ﺫ 󰆳󰆎  ،ﺹﺍ  ﺍﻭ 󰄨  ًﺃ 󰉼󰉫 ﺓﺀﺍﺍ ﺓﺩ ﺹ  ﻁ ،ﺍ ﺍ ﻭﺃ ﺩ󰁅ﺍ ﺕﺍﺭﺍﻭ ،ﺕﺍ ﻭﺃ .ﺍﺍ ﺕﺍﻭﺩﺃ 󰆭󰆠ﺇ ﻝﺍﻭ ﺕﺍ 󰊜󰊈ﺫ 󰆳󰆎  ،ً ﺹﺍ ﻉﺍﺃ 󰄨󰄚  ﻯﺍ ﺍ 󰆳󰆎 󰋜󰊈  – C2 ًﺃ 󰉼󰉫 .ﺍ ﺓ ﻭﺃ ﺍ   ﺓ ﻭﺃ ﺓﺩﻥ ﺍ ﺩﺍ ﻭ ﺩﺍ 󰆳󰅤ﺍﻭﺏﺍ 󰆳󰆎 󰃕󰃀ﺍ ﺕﻭﺍ ﻭ،ﺓﺍﻭ󰃎󰃆ﺍﺹﺍ  ﺓ 󰄨 .ﺍﻭ ﺍ Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 9751 CEFR specifications for reading comprehension in Hindi (hi) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Hindi text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 -       ,           ,  ,                A2 -          ,     ,         B1 -                         B2 -              , -               ,                  ,               C1 -      ,      ,          ,        ,     ,           -     ,               C2 -             , ,          -              ,                Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 3 9752 CEFR specifications for reading comprehension in Russian (ru) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Russian text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 — учащиеся этого уровня могут понимать очень короткие, простые тексты по одной фразе за раз, подбирая знакомые имена, слова и основные фразы и перечитывая их по мере необходимости. A2 — учащиеся этого уровня могут понимать короткие, простые тексты, содержащие наиболее частотную лексику, включая часть общих международных словарных единиц. B1 — учащиеся этого уровня могут читать простые фактические тексты по темам, связанным с их областью интересов, с удовлетворительным уровнем понимания. B2 — учащиеся этого уровня могут читать с большой степенью независимости, адаптируя стиль и скорость чтения к различным текстам и целям и выборочно используя соответствующие справочные источники. Имеет широкий активный словарный запас чтения, но может испытывать некоторые трудности с редко встречающимися идиомами. C1 — учащиеся этого уровня могут понимать в деталях длинные, сложные тексты, независимо от того, относятся ли они к их собственной области специализации, при условии, что они могут перечитывать сложные разделы. Они также могут понимать широкий спектр текстов, включая литературные произведения, газетные или журнальные статьи, а также специализированные академические или профессиональные публикации, при условии, что есть возможности для перечитывания и у них есть доступ к справочным материалам. C2 - Учащиеся этого уровня могут понимать практически все типы текстов, включая абстрактные, структурно сложные или очень разговорные литературные и нелитературные произведения. Они также могут понимать широкий спектр длинных и сложных текстов, оценивая тонкие различия стиля и неявного, а также явного значения. Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 4 9753 CEFR specifications for reading comprehension in Welsh (CY) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Welsh text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 - Gall dysgwyr y lefel hon ddeall testunau byr iawn, syml un cymal ar y tro, gan godi enwau, geiriau ac ymadroddion sylfaenol cyfarwydd ac ailddarllen yn ôl yr angen. A2 - Gall dysgwyr y lefel hon ddeall testunau byr, syml sy’n cynnwys yr eirfa fwyaf aml, gan gynnwys cyfran o eitemau geirfa ryngwladol a rennir. B1 - Gall dysgwyr y lefel hon ddarllen testunau ffeithiol syml ar bynciau sy’n ymwneud â’u maes diddordeb gyda lefel foddhaol o ddealltwriaeth. B2 - Gall dysgwyr y lefel hon ddarllen yn annibynnol iawn, gan addasu arddull a chyflymder darllen i wahanol destunau a dibenion, a defnyddio ffynonellau cyfeirio priodol yn ddetholus. Yn meddu ar eirfa ddarllen weithredol eang, ond gall brofi peth anhawster gydag idiomau amledd isel. C1 - Gall dysgwyr y lefel hon ddeall yn fanwl destunau hir a chymhleth, p’un a yw’r rhain yn ymwneud â’u maes arbenigedd eu hunain ai peidio, ar yr amod eu bod yn gallu ailddarllen adrannau anodd. Gallant hefyd ddeall amrywiaeth eang o destunau gan gynnwys ysgrifau llenyddol, erthyglau papur newydd neu gylchgronau, a chyhoeddiadau academaidd neu broffesiynol arbenigol, ar yr amod bod cyfleoedd i’w hail-ddarllen a bod offer cyfeirio ar gael iddynt. C2 - Gall dysgwyr ar y lefel hon ddeall bron bob math o destunau gan gynnwys ysgrifau llenyddol ac anllenyddol haniaethol, strwythurol gymhleth, neu ysgrifau llenyddol ac anllenyddol hynod lafar. Gallant hefyd ddeall ystod eang o destunau hir a chymhleth, gan werthfawrogi gwahaniaethau cynnil o ran arddull ac ystyr ymhlyg yn ogystal ag ystyr amlwg. Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 9754 CEFR specifications for written production in English (EN) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 - Learners of this level can give information about matters of personal relevance (e.g. likes and dislikes, family, pets) using simple words/signs and basic expressions. Learners can also produce simple isolated phrases and sentences. A2 - Learners of this level can produce a series of simple phrases and sentences linked with simple connectors like “and”, “but” and “because”. Learners have sufficient vocabulary for the expression of basic communicative needs and for coping with simple survival needs. B1 - Learners of this level can produce straightforward connected texts on a range of familiar subjects within their field of interest, by linking a series of shorter discrete elements into a linear sequence. Learners have a good range of vocabulary related to familiar topics and everyday situations. B2 - Learners of this level can produce clear, detailed texts on a variety of subjects related to their field of interest, synthesising and evaluating information and arguments from a number of sources. Learners have a good range of vocabulary for matters connected to their field and most general topics. C1 - Learners of this level can produce clear, well-structured texts of complex subjects, underlining the relevant salient issues, expanding and supporting points of view at some length with subsidiary points, reasons and relevant examples, and rounding off with an appropriate conclusion. Learners can alsoemploy the structure and conventions of a variety of genres, varying the tone, style and register according to addressee, text type and theme. C2 - Learners of this level can produce clear, smoothly flowing, complex texts in an appropriate and effective style and a logical structure which helps the reader identify significant points. Learners have a good command of a very broad lexical repertoire including idiomatic expressions and colloquialisms; shows awareness of connotative levels of meaning. Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 9755 CEFR specifications for written production in Estonian (ET) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Estonian text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 – Selle taseme õppijad suudavad anda teavet isiklikult olulistel teemadel (nt meeldimised ja mittemeeldimised, perekond, lemmikloomad), kasutades lihtsaid sõnu/viipeid ja põhilisi väljendeid. Õppijad suudavad moodustada ka lihtsaid üksikuid fraase ja lauseid. A2 – Selle taseme õppijad suudavad toota lihtsate fraaside ja lausete jada, mis on seotud lihtsate sidesõnadega nagu „ja“, „aga“ ja „sest“. Neil on piisav sõnavara põhiliste suhtlusvajaduste ja lihtsate ellujäämisvajaduste rahuldamiseks. B1 – Selle taseme õppijad suudavad koostada arusaadavaid, seotud tekste tuttavatel teemadel oma huvivaldkonnas, sidudes lühemaid üksikuid elemente lineaarseks järjestuseks. Neil on hea sõnavara tuttavate teemade ja igapäevaste olukordade kirjeldamiseks. B2 – Selle taseme õppijad suudavad koostada selgeid ja üksikasjalikke tekste erinevatel nende huvivaldkonnaga seotud teemadel, sünteesides ja hinnates teavet ja argumente mitmest allikast. Neil on hea sõnavara oma valdkonnaga seotud teemadeks ning enamike üldiste teemade jaoks. C1 – Selle taseme õppijad suudavad koostada selgeid ja hästi struktureeritud tekste keerukatel teemadel, tuues esile olulised küsimused, laiendades ja toetades seisukohti üksikasjalikult koos täiendavate punktide, põhjuste ja asjakohaste näidetega ning lõpetades sobiva järeldusega. Samuti suudavad nad kasutada erinevate žanrite struktuuri ja konventsioone ning varieerida tooni, stiili ja registrit vastavalt adressaadile, tekstiliigile ja teemale. C2 – Selle taseme õppijad suudavad koostada selgeid, sujuvaid ja keerukaid tekste sobivas ja tõhusas stiilis ning loogilises struktuuris, mis aitab lugejal tuvastada olulisi punkte. Neil on väga lai sõnavara, mis sisaldab idioome ja kõnekeelseid väljendeid; nad tunnetavad ka tähenduse konnotatiivseid tasandeid. Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 9762 CEFR specifications for written production in Portuguese (PT) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Portuguese text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 – Os aprendentes deste nível conseguem fornecer informações sobre assuntos de relevância pessoal (por exemplo, gostos e preferências, família, animais de estimação) usando palavras/sinais simples e expressões básicas. Também conseguem produzir frases e expressões simples e isoladas. A2 – Os aprendentes deste nível conseguem produzir uma série de frases e expressões simples ligadas por conectores básicos como “e”, “mas” e “porque”. Têm vocabulário suficiente para expressar necessidades comunicativas básicas e lidar com necessidades simples de sobrevivência. B1 – Os aprendentes deste nível conseguem produzir textos simples e coerentes sobre uma variedade de temas familiares dentro de seu campo de interesse, ligando uma série de elementos mais curtos em sequência linear. Possuem um bom repertório de vocabulário relacionado a temas familiares e situações do cotidiano. B2 – Os aprendentes deste nível conseguem produzir textos claros e detalhados sobre uma variedade de assuntos relacionados ao seu campo de interesse, sintetizando e avaliando informações e argumentos de várias fontes. Têm um bom vocabulário para assuntos relacionados à sua área e à maioria dos temas gerais. C1 – Os aprendentes deste nível conseguem produzir textos claros e bem estruturados sobre temas complexos, destacando os pontos relevantes, desenvolvendo e sustentando pontos de vista com argumentos secundários, razões e exemplos pertinentes, e encerrando com uma conclusão apropriada. Também conseguem empregar a estrutura e as convenções de diferentes gêneros textuais, variando o tom, o estilo e o registro conforme o destinatário, o tipo de texto e o tema. C2 – Os aprendentes deste nível conseguem produzir textos claros, fluidos e complexos em um estilo apropriado e eficaz, com uma estrutura lógica que ajuda o leitor a identificar os pontos significativos. Têm um excelente domínio de um repertório lexical muito amplo, incluindo expressões idiomáticas e coloquialismos, e demonstram consciência dos níveis conotativos de significado. Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 9763 CEFR specifications for written production in Arabic (ar) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Arabic text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: :) 󰃎󰃆 ﺕﺍﺫ ﺍ ﻝ ﺕ 󰇶󰇚 ﻯﺍ ﺍ 󰆳󰆎 󰋜󰊈  - A1 ﺕﺍﻭ  ﺕﺍﺭﺇ/ﺕ󰀖 ﻡﺍ (ﺍ ﺕﺍﺍ ،󰃎󰃆ﺍ ، ﻭ   .ﺓﺩﻭ  󰄨󰄚ﻭ ﺕﺍﺭ ﺝﺇ ًﺃ 󰉼󰉫 󰁿󰁱󰁑 .ﺃ ﺍﺍ ﺍ ﺍﻭ ﺕﺍﺭﺍ  󰃎󰃆 ﺝﺇ ﻯﺍ ﺍ 󰆳󰆎 󰋜󰊈  - A2  󰊎󰊈 󰁅 ﺕﺍﺩ 󰉼󰈸󰃕󰃀 ."ﻥ"ﻭ ،"ـ" ،"ﻭ"   ﺭ ﺕﺍﻭﺩﺃ ﻡﺍ .ﺍ ﺀﺍ ﺕﺍ  ﺍﻭ ﺍ ﺍﺍ ﺕﺍ ﺍﺍ  󰄨 ﻝ ﻭ ﺍ ﺹ ﺝﺇ ﻯﺍ ﺍ 󰆳󰆎 󰋜󰊈  - B1  󰆳󰆎 ﺍ 󰃎󰃆ﺍ ﺍ  󰃎󰃆 ﺭ   ،󰉼󰓜ﺍ ﻝ󰆳󰆎 ﺍ .ﺍ ﺍﺍﻭ ﺍ ﺍ ﺍ ﺕﺍﺩﺍ  ﺓ 󰄨 󰉼󰈸󰃕󰃀 .   󰄨 ﻝ 󰃎󰃆ﻭ ﺍﻭ ﺹ ﺝﺇ ﻯﺍ ﺍ 󰆳󰆎 󰋜󰊈  - B2 .ﺭﺩﺍ  ﺩ  󰌔󰋿ﺍﻭ ﺕﺍ ﻭ   ،󰉼󰓜ﺍ ﻝ ﺍ ﺍﺍ .ﺍ ﺍﺍ ﻭ  ﺍ ﺭ ﺍ ﺕﺍﺩﺍ  ﺓ 󰄨 󰉼󰈸󰃕󰃀 ﺍ ﻝ  󰁀󰀞 ﻭ ﺍﻭ ﺹ ﺝﺇ ﻯﺍ ﺍ 󰆳󰆎 󰋜󰊈  - C1 ﺏﺃﻭ  ﺭ󰁅󰀞 ﺍ ﻁﺍ ﻭ ،󰃎󰃆ﺍ ﺕﺍﺫ ﺓﺯﺭﺍ ﺍ ﺯﺍﺇ  ،ﺓ   󰄨 ﺕﺍﻭ  ﻡﺍﺍ 󰉼󰉫 󰁿󰁱󰁑 .  ﺀ󰉼󰉮󰙗ﺍﻭ ، 󰃎󰃆ﺃﻭ .ﻉﺍﻭ ﺍ ﻉﻭ ﺍ  ﺍﻭ ﺏﺍﻭ ﺍ ﻭ ،ﺍ ﻁﺍ  ﺏ ﺓﻭ ،ﻭ ،ﺍﻭ ﺹ ﺝﺇ ﻯﺍ ﺍ 󰆳󰆎 󰋜󰊈  - C2 ﺍﻭ 󰄨 󰆳󰆎   󰉼󰈸󰃕󰃀 .ﺍ ﻁﺍ   ﺉﺭﺍ   ﻭ ﻝﻭ ﺕ ًﻭ ﻥﻭُﻭ ؛ﺍ ﺕ󰀖󰀞ﺍﻭ ﺍ ﺍ  ﺕﺍﺩﺍ  ﺍً .ﺍ ﺍ Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 9764 CEFR specifications for written production in Hindi (hi) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Hindi text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 -          (  , ,  )    /             , -          A2 -               ,  “”, “”  “ ”                       B1 -       ,        ,        , - -                       B2 -             ,     ,                           C1 -        ,       ,        ,    ,           ,                   ,  ,        ,         C2 -      ,                                         ,   -       ;    (connotative)       Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 6 9765 CEFR specifications for written production in Russian (ru) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Russian text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 — Обучающиеся на этом уровне могут сообщать информацию на личные темы (например, о своих предпочтениях, семье, домашних животных), используя простые слова/жесты и базовые выражения. Также они могут составлять простые отдельные фразы и предложения. A2 — Обучающиеся на этом уровне могут строить серию простых фраз и предложений, соединённых с помощью простых союзов, таких как «и», «но» и «потому что». У них есть достаточный словарный запас для выражения базовых коммуникативных потребностей и для решения простых бытовых задач. B1 — Обучающиеся на этом уровне могут создавать понятные связные тексты на знакомые темы в рамках своей области интересов, объединяя серию коротких, отдельных элементов в линейную последовательность. У них хороший запас слов, связанных с повседневными ситуациями и знакомыми темами. B2 — Обучающиеся на этом уровне могут писать чёткие и подробные тексты по различным темам, связанным с их сферой интересов, обобщая и оценивая информацию и аргументы из нескольких источников. У них хороший словарный запас по тематике своей области и большинству общих тем. C1 — Обучающиеся на этом уровне могут создавать чёткие, хорошо структурированные тексты на сложные темы, подчёркивая важные аспекты, развивая и обосновывая свою точку зрения с помощью дополнительных аргументов, причин и релевантных примеров, и завершать текст уместным заключением. Они также могут применять структуру и нормы различных жанров, варьируя тон, стиль и регистр в зависимости от адресата, типа текста и темы. C2 - Обучающиеся на этом уровне могут создавать чёткие, плавные и сложные тексты в уместном и эффективном стиле, с логичной структурой, которая помогает читателю выделять важные моменты. У них отличное владение очень широким лексическим запасом, включая идиоматические выражения и разговорную лексику; они осознают коннотативные уровни значений. Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 7 9766 CEFR specifications for written production in Welsh (CY) You are an expert in language proficiency classification based on the Common European Framework of Reference for Languages (CEFR). Your task is to analyze the given Welsh text or narrative and determine the best CEFR level [A1, A2, B1, B2, C1, or C2] based on the CEFR descriptors of reading comprehension of learners below: A1 – Gall dysgwyr ar y lefel hon roi gwybodaeth am faterion o berthnasedd personol (e.e. pethau maen nhw’n eu hoffi a’u casáu, teulu, anifeiliaid anwes) gan ddefnyddio geiriau/arwyddion syml ac ymadroddion sylfaenol. Gall dysgwyr hefyd gynhyrchu brawddegau ac ymadroddion syml, arwahanol. A2 – Gall dysgwyr ar y lefel hon gynhyrchu cyfres o ymadroddion a brawddegau syml wedi’u cysylltu gan gysyllteiriau syml fel “a”, “ond” a “oherwydd”. Mae gan ddysgwyr eirfa ddigonol i fynegi anghenion cyfathrebu sylfaenol ac i ymdopi ag anghenion goroesi syml. B1 – Gall dysgwyr ar y lefel hon gynhyrchu testunau cysylltiedig, uniongyrchol ar ystod o bynciau cyfarwydd o fewn eu maes diddordeb, drwy gysylltu cyfres o elfennau byrrach ar wahân i mewn i ddilyniannol linol. Mae ganddynt ystod dda o eirfa sy’n ymwneud â phethau cyfarwydd a sefyllfaoedd bob dydd. B2 – Gall dysgwyr ar y lefel hon gynhyrchu testunau clir, manwl ar amrywiaeth o bynciau sy’n gysylltiedig â’u maes diddordeb, gan gyfuno a gwerthuso gwybodaeth a dadleuon o sawl ffynhonnell. Mae ganddynt ystod dda o eirfa ar gyfer materion sy’n gysylltiedig â’u maes ac ar gyfer y rhan fwyaf o bynciau cyffredinol. C1 – Gall dysgwyr ar y lefel hon gynhyrchu testunau clir, wedi’u strwythuro’n dda ar bynciau cymhleth, gan amlygu’r materion perthnasol, ehangu a chefnogi safbwyntiau’n fanwl gyda phwyntiau ategol, rhesymau ac enghreifftiau perthnasol, a gorffen gyda chasgliad priodol. Gallant hefyd ddefnyddio strwythur a chonfensiynau amrywiaeth o genres, gan amrywio’r naws, arddull a chofrestr yn ôl y derbynnydd, math y testun a’r thema. C2 – Gall dysgwyr ar y lefel hon gynhyrchu testunau clir, esmwyth a chymhleth mewn arddull briodol ac effeithiol ac mewn strwythur resymegol sy’n helpu’r darllenydd i nodi pwyntiau arwyddocaol. Mae ganddynt reolaeth dda dros eirfa eang iawn gan gynnwys ymadroddion idiomatig a llafariad; maent yn dangos ymwybyddiaeth o lefelau ystyron cynhennus. Provide only the CEFR level as output directly, without explanation or justification. Text: «TEXT» Answer: 9767