Open-weight LLMs have been released by frontier labs; however, sovereign Large Language Models (for languages other than English) remain low in supply yet high in demand. Training large language models (LLMs) for low-resource languages such as Hebrew poses unique challenges. In this paper, we introduce Dicta-LM 3.0: an open-weight collection of LLMs trained on substantially-sized corpora of Hebrew and English texts. The model is released in three sizes: 24B - adapted from the Mistral-Small-3.1 base model, 12B - adapted from the NVIDIA Nemotron Nano V2 model, and 1.7B - adapted from the Qwen3-1.7B base model. We are releasing multiple variants of each model, each with a native context length of 65k tokens; base model and chat model with tool-calling support. To rigorously evaluate our models, we introduce a new benchmark suite for evaluation of Hebrew chat-LLMs, covering a diverse set of tasks including Translation, Summarization, Winograd, Israeli Trivia, and Diacritization (nikud). Our work not only addresses the intricacies of training LLMs in low-resource languages but also proposes a framework that can be leveraged for adapting other LLMs to various non-English languages, contributing to the broader field of multilingual NLP.

Since their initial release, BERT models have demonstrated exceptional performance on a variety of tasks, despite their relatively small size (BERT-base has 100M parameters). Nevertheless, the architectural choices used in these models are outdated compared to newer transformer-based models such as Llama3 and Qwen3. In recent months, several architectures have been proposed to close this gap. ModernBERT and NeoBERT both show strong improvements on English benchmarks and significantly extend the supported context window. Following their successes, we introduce NeoDictaBERT and NeoDictaBERT-bilingual: BERT-style models trained using the same architecture as NeoBERT, with a dedicated focus on Hebrew texts. These models outperform existing ones on almost all Hebrew benchmarks and provide a strong foundation for downstream tasks. Notably, the NeoDictaBERT-bilingual model shows strong results on retrieval tasks, outperforming other multilingual models of similar size. In this paper, we describe the training process and report results across various benchmarks. We release the models to the community as part of our goal to advance research and development in Hebrew NLP.

Training large language models (LLMs) in low-resource languages such as Hebrew poses unique challenges. In this paper, we introduce DictaLM2.0 and DictaLM2.0-Instruct, two LLMs derived from the Mistral model, trained on a substantial corpus of approximately 200 billion tokens in both Hebrew and English. Adapting a pre-trained model to a new language involves specialized techniques that differ significantly from training a model from scratch or further training existing models on well-resourced languages such as English. We outline these novel training methodologies, which facilitate effective learning and adaptation to the linguistic properties of Hebrew. Additionally, we fine-tuned DictaLM2.0-Instruct on a comprehensive instruct dataset to enhance its performance on task-specific instructions. To rigorously evaluate our models, we introduce a new benchmark suite for Hebrew LLM evaluation, covering a diverse set of tasks including Question Answering, Sentiment Analysis, Winograd Schema Challenge, Translation, and Summarization. Our work not only addresses the intricacies of training LLMs in low-resource languages but also proposes a framework that can be leveraged for adapting other LLMs to various non-English languages, contributing to the broader field of multilingual NLP.

Syntactic parsing remains a critical tool for relation extraction and information extraction, especially in resource-scarce languages where LLMs are lacking. Yet in morphologically rich languages (MRLs), where parsers need to identify multiple lexical units in each token, existing systems suffer in latency and setup complexity. Some use a pipeline to peel away the layers: first segmentation, then morphology tagging, and then syntax parsing; however, errors in earlier layers are then propagated forward. Others use a joint architecture to evaluate all permutations at once; while this improves accuracy, it is notoriously slow. In contrast, and taking Hebrew as a test case, we present a new "flipped pipeline": decisions are made directly on the whole-token units by expert classifiers, each one dedicated to one specific task. The classifiers are independent of one another, and only at the end do we synthesize their predictions. This blazingly fast approach sets a new SOTA in Hebrew POS tagging and dependency parsing, while also reaching near-SOTA performance on other Hebrew NLP tasks. Because our architecture does not rely on any language-specific resources, it can serve as a model to develop similar parsers for other MRLs.

We present DictaBERT, a new state-of-the-art pre-trained BERT model for modern Hebrew, outperforming existing models on most benchmarks. Additionally, we release three fine-tuned versions of the model, designed to perform three specific foundational tasks in the analysis of Hebrew texts: prefix segmentation, morphological tagging and question answering. These fine-tuned models allow any developer to perform prefix segmentation, morphological tagging and question answering of a Hebrew input with a single call to a HuggingFace model, without the need to integrate any additional libraries or code. In this paper we describe the details of the training as well and the results on the different benchmarks. We release the models to the community, along with sample code demonstrating their use. We release these models as part of our goal to help further research and development in Hebrew NLP.

We present DictaLM, a large-scale language model tailored for Modern Hebrew. Boasting 7B parameters, this model is predominantly trained on Hebrew-centric data. As a commitment to promoting research and development in the Hebrew language, we release both the foundation model and the instruct-tuned model under a Creative Commons license. Concurrently, we introduce DictaLM-Rab, another foundation model geared towards Rabbinic/Historical Hebrew. These foundation models serve as ideal starting points for fine-tuning various Hebrew-specific tasks, such as instruction, Q&A, sentiment analysis, and more. This release represents a preliminary step, offering an initial Hebrew LLM model for the Hebrew NLP community to experiment with.

Semitic morphologically-rich languages (MRLs) are characterized by extreme word ambiguity. Because most vowels are omitted in standard texts, many of the words are homographs with multiple possible analyses, each with a different pronunciation and different morphosyntactic properties. This ambiguity goes beyond word-sense disambiguation (WSD), and may include token segmentation into multiple word units. Previous research on MRLs claimed that standardly trained pre-trained language models (PLMs) based on word-pieces may not sufficiently capture the internal structure of such tokens in order to distinguish between these analyses.Taking Hebrew as a case study, we investigate the extent to which Hebrew homographs can be disambiguated and analyzed using PLMs. We evaluate all existing models for contextualized Hebrew embeddings on a novel Hebrew homograph challenge sets that we deliver. Our empirical results demonstrate that contemporary Hebrew contextualized embeddings outperform non-contextualized embeddings; and that they are most effective for disambiguating segmentation and morphosyntactic features, less so regarding pure word-sense disambiguation. We show that these embeddings are more effective when the number of word-piece splits is limited, and they are more effective for 2-way and 3-way ambiguities than for 4-way ambiguity. We show that the embeddings are equally effective for homographs of both balanced and skewed distributions, whether calculated as masked or unmasked tokens. Finally, we show that these embeddings are as effective for homograph disambiguation with extensive supervised training as with a few-shot setup.

We present a new pre-trained language model (PLM) for Rabbinic Hebrew, termed Berel (BERT Embeddings for Rabbinic-Encoded Language). Whilst other PLMs exist for processing Hebrew texts (e.g., HeBERT, AlephBert), they are all trained on modern Hebrew texts, which diverges substantially from Rabbinic Hebrew in terms of its lexicographical, morphological, syntactic and orthographic norms. We demonstrate the superiority of Berel on Rabbinic texts via a challenge set of Hebrew homographs. We release the new model and homograph challenge set for unrestricted use.

We present a new pre-trained language model (PLM) for modern Hebrew, termed AlephBERTGimmel, which employs a much larger vocabulary (128K items) than standard Hebrew PLMs before. We perform a contrastive analysis of this model against all previous Hebrew PLMs (mBERT, heBERT, AlephBERT) and assess the effects of larger vocabularies on task performance. Our experiments show that larger vocabularies lead to fewer splits, and that reducing splits is better for model performance, across different tasks. All in all this new model achieves new SOTA on all available Hebrew benchmarks, including Morphological Segmentation, POS Tagging, Full Morphological Analysis, NER, and Sentiment Analysis. Subsequently we advocate for PLMs that are larger not only in terms of number of layers or training data, but also in terms of their vocabulary. We release the new model publicly for unrestricted use.

Many classical texts are available in multiple versions that almost always differ from each other due to transcription error and editorial discretion. One of the central challenges in the study of such texts is the preparation of a ‘synoptic’ text: an aligned presentation of the various versions in which corresponding words or phrases, even if not identical, are mapped to each other. Multiple text alignment of this sort must take into account orthographic and conceptual relationships between words. In this article, we define this text alignment problem as an optimization problem by providing a formal measure of alignment quality. Unlike previous measures, our measure uses word embeddings to take into account conceptual similarity between aligned words. We propose an efficient and scalable alignment method in accordance with the proposed criteria. This method splits the texts to be aligned into smaller subtexts, thus improving both efficiency and accuracy. Empirical comparisons on sample data indicate our method is significantly faster than existing methods, often rendering intractable problems tractable, and that the alignment obtained by our method is considerably better than that obtained by other methods.

One of the primary tasks of morphological parsers is the disambiguation of homographs. Particularly difficult are cases of unbalanced ambiguity, where one of the possible analyses is far more frequent than the others. In such cases, there may not exist sufficient examples of the minority analyses in order to properly evaluate performance, nor to train effective classifiers. In this paper we address the issue of unbalanced morphological ambiguities in Hebrew. We offer a challenge set for Hebrew homographs -- the first of its kind -- containing substantial attestation of each analysis of 21 Hebrew homographs. We show that the current SOTA of Hebrew disambiguation performs poorly on cases of unbalanced ambiguity. Leveraging our new dataset, we achieve a new state-of-the-art for all 21 words, improving the overall average F1 score from 0.67 to 0.95. Our resulting annotated datasets are made publicly available for further research.

We consider a thought experiment in which voters could submit binary preferences regarding each of a pre-determined list of independent relevant issues, so that majorities could be tallied per issue. It might be thought that if such voting became technically feasible and widespread, parties and coalitions could be circumvented altogether and would become irrelevant. In this paper, we show, however, why and how voters would spontaneously self-organize into parties, and parties would self-organize into coalitions, prior to elections. We will see that such coordination is possible, even assuming very limited capabilities of communication and coordination. Using both analytical and empirical methods, we show that the average voter in a majority coalition would gain more than if no parties were formed, but the average voter overall (in or out of the coalition) would be worse off. Furthermore, the extent of these gains and losses is inversely proportional to the degree to which voters line along a unidimensional left–right axis.

We present a system for automatic diacritization of Hebrew text. The system combines modern neural models with carefully curated declarative linguistic knowledge and comprehensive manually constructed tables and dictionaries. Besides providing state of the art diacritization accuracy, the system also supports an interface for manual editing and correction of the automatic output, and has several features which make it particularly useful for preparation of scientific editions of Hebrew texts. The system supports Modern Hebrew, Rabbinic Hebrew and Poetic Hebrew. The system is freely accessible for all use at http://nakdanpro.dicta.org.il.

Brain-computer interfaces (BCIs) have been employed to provide different patient groups with communication and control that does not require the use of limbs that have been damaged. In this study, we explored BCI-based navigation in three long term amputees. Each participant attempted motor execution with the affected limb, and performed motor execution with the intact limb, while fMRI activity was recorded. Participants attempted, and executed, one of four tasks to direct the movement of an avatar on a monitor. Classification accuracy was very high across both cue-based and free-choice conditions. Results support the use of this fMRI BCI approach for virtual navigation, which could improve BCIs based on fMRI as well as other approaches such as EEG.

Purpose: The purpose of this study is to investigate the relationship between production and salary and rewards of senior employees.Design/methodology/approach: The Support Vector Machines (SVM) machine-learning algorithm classifier, was used to distinguish between production and non-production sentiments.Findings: Senior managers sent significantly more work-related emails than non-senior employees. However, the length of work-related emails was found to be significantly greater among non-seniors. Finally, the gap between the two groups of workers in terms of productivity is minor, especially as compared to the gap in salaries. Originality: This study offers an efficient tool to automate employee productivity assessment and use it to prove the unjustified compensations at Enron Corporation, an energy company that went bankrupt

We have developed Brain-Voyant, an efficient general-purpose machine learning tool for real-time functional magnetic resonance imaging classification using whole-brain data, which can be used to explore novel brain-computer interface paradigms or advanced neurofeedback protocols. We have created a convenient and configurable front-end tool that receives fMRI-based multi-voxel raw brain data as input. Our tool processes, analyses, classifies and transfers the classification to an external object such as a virtual avatar or a humanoid robot in real-time. Our tool is focused on minimizing delay time, and to that end, it employs a method that is based on examining in advance the voxels that have been found to be task-relevant in the machine learning model training phase.The tool's code base was designed to be easily extended to support additional feature reduction, normalization and classification algorithms. This tool was used in several published studies using motor execution, motor imagery, and visual category classification in cue-based and free-choice brain-computer interface experiments, with both healthy and amputated subjects. This tool is not limited by number of classes, is not limited to predefined regions of interest, and classifier instances can run in parallel to combine multiple classification tasks in real time. Finally, our tool is able use the slow peaking blood-oxygen-level dependent signal to classify our subjects' intention during the two-second window TR. We release this tool as open-source for non-commercial usage.

The ability to allow subjects, including paralyzed patients, to perform a task using brain-computer interfaces has seen a rapid and growing success. Surprisingly, however, it is still not known how far such performance can be improved - especially in cases of long term amputation where both efferent and afferent functions are abolished and may lead to deterioration of the relevant brain representations. Here we used real-time fMRI to demonstrate a remarkably high performance of long term amputees in controlling a computer generated avatar using their missing hand. The missing limb BCI performance showed similar levels both when compared to the intact hand and to control participants.

This paper demonstrates the use of genetic algorithms for evolving a grandmaster-level evaluation function for a chess program. This is achieved by combining supervised and unsupervised learning. In the supervised learning phase the organisms are evolved to mimic the behavior of human grandmasters, and in the unsupervised learning phase these evolved organisms are further improved upon by means of coevolution. While past attempts succeeded in creating a grandmaster-level program by mimicking the behavior of existing computer chess programs, this paper presents the first successful attempt at evolving a state-of-the-art evaluation function by learning only from databases of games played by humans. Our results demonstrate that the evolved program outperforms a two-time World Computer Chess Champion.

: In previous studies we have shown that real-time fMRI, despite the low temporal resolution of the brain oxygen level dependent (BOLD) signal, can be used as a brain-computer interface for navigation tasks, using motor imagery and motor execution. Here we describe a pilot study aimed at leveraging the superior spatial resolution of fMRI and describe a BCI paradigm going beyond a single brain network for control, retaining an intuitive mapping between brain activity and BCI functionality. The experiment simulates non-trivial navigation and item selection tasks. We allow a subject to teleoper-ate an HRP-4 humanoid robot: motor actions are mapped into simple navigation commands to navigate the robot inside a room and visual attention is mapped to direct the robot’s arm toward one of three objects placed on a table. When the correct item has been selected, the subject navi-gates the robot toward the experimenter in order to simulate the delivery of the object. Here we describe the method based on seven classes, of-fline classification results, and results from a real time pilot with a single subject performing the task several times.