Our Work

A performance-portable C++/Kokkos port of the ocean and sea-ice core of FESOM2: one source tree for CPUs (Kokkos Serial, OpenMP) and GPUs (CUDA, HIP). The Serial build reproduces the C port byte for b…

A performance-portable C++/Kokkos port of the ocean and sea-ice core of FESOM2: one source tree for CPUs (Kokkos Serial, OpenMP) and GPUs (CUDA, HIP). The Serial build reproduces the C port byte for byte. Code of the paper An Ocean Model Ported by a Large Language Model: Experience and Lessons from FESOM2 (Fortran to C to C++/Kokkos). Developed at github.com/koldunovn/fesom_kokkos.Files:fesom_kokkos-1.0.0.zip: version 1.0.0. Start with docs/RUNNING.md.fesom_kokkos-0bd65156.zip: commit 0bd6515, the paper's CUDA hindcast 1958-2020.fesom_kokkos-a0b474b8.zip: commit a0b474b8, the one-year byte-identity test and the GPU acceptance tests.fesom_kokkos-4436583b.zip: commit 4436583b, the CPU strong-scaling measurements on Levante.fesom_kokkos-d1a56907.zip: commit d1a5690, the GPU strong-scaling measurements on Levante, LUMI and MareNostrum 5.kokkos-4.4.01.zip: Kokkos 4.4.01 (Apache License 2.0 with LLVM exceptions), the git submodule the model builds in-tree; unpack it into externals/kokkos.Please also cite FESOM2 (Danilov et al., 2017, doi:10.5194/gmd-10-765-2017) and Kokkos (Trott et al., 2022, doi:10.1109/TPDS.2021.3097283).

The two versions of the FESOM2 ocean and sea-ice model used as the Fortran reference in the paper An Ocean Model Ported by a Large Language Model: Experience and Lessons from FESOM2 (Fortran to C to C…

The two versions of the FESOM2 ocean and sea-ice model used as the Fortran reference in the paper An Ocean Model Ported by a Large Language Model: Experience and Lessons from FESOM2 (Fortran to C to C++/Kokkos), from github.com/FESOM/fesom2.fesom2-llmport-fortran-reference-9271ae92.zip: tag llmport-fortran-reference (commit 9271ae92), the tree both ports were translated from: upstream d729beaa plus the cold-start wind-rotation fix and the reference-dump instrumentation used for the per-kernel comparisons.fesom2-llmport-fortran-reference-detfill-e63430b3.zip: tag llmport-fortran-reference-detfill (commit e63430b3), the same tree plus the deterministic initial-condition fill (namelist switch ic_extrap_det), as compiled for the 1958-2020 reference hindcast. A later revision of this change was merged into FESOM2 as pull request 979.Built with Intel oneAPI 2022.0.1 and Open MPI 4.1.2 on Levante (DKRZ).

The reduced data behind every figure and table of the paper An Ocean Model Ported by a Large Language Model: Experience and Lessons from FESOM2 (Fortran to C to C++/Kokkos), the scripts that produce t…

The reduced data behind every figure and table of the paper An Ocean Model Ported by a Large Language Model: Experience and Lessons from FESOM2 (Fortran to C to C++/Kokkos), the scripts that produce the figures and tables from them, the job scripts of the model runs, and the records that tie every number to the binary that produced it (data/build_provenance.json). ARCHIVE_README.md describes the contents, how to rebuild the figures, which code produced which result, and the limitations. The archive was verified by rebuilding every figure from it alone. The model output itself (about 1 TB) is not included.

Cardiovascular Disease(CVD) is the primary cause of morbidity and mortality in Type2 Diabetes globally due to interactions between vascular and metabolic mechanisms. In these patients, disease progres…

Cardiovascular Disease(CVD) is the primary cause of morbidity and mortality in Type2 Diabetes globally due to interactions between vascular and metabolic mechanisms. In these patients, disease progression is more complex, making early prediction challenging. This research implements an Integrated 1DCNN-BiLSTM model for accurate cardiac disease prediction in diabetic individuals.It comprises a 1-dimensional convolutional neural network to detect the spatial correlations of the data combine with Bidirectional Long Short-Term Memory (BiLSTM) that identify the forward and backward dependencies for an efficient analysis of sequential clinical data. The proposed model avoids the limitations of the Framingham Risk Score, which relies on manual computation and a limited set of features. The primary aim is to improve predictive accuracy while offering an economical and flexible solution for early detection, especially in resource-limited settings. Firstly,the model separate the diabetic data from Cleveland–Hungarian cardiac data sets and perform with an accuracy of 96 % for the prediction of cardiac diseases in diabetic patients. Additionally ,the model perform well over traditional methods: Support Vector Machine, Logistic Regression, Decision Tree, Naive Bayes, Random Forest and conventional 1DCNN model .The Novelty lies in transforming tabular clinical data into a sequential format to exploit temporal dependencies using BiLSTM, an approach not commonly used in traditional CVD prediction methods. Furthermore , the integrated Recursive Feature Elimination (RFE) and Principal Component Analysis (PCA) during pre-processing enhance its prediction accuracy and efficiency . Another significant contribution is that the model is applied on the NVIDIA Jetson Nano, enabling low-cost, edge-based inference for real-time IoT-enabled healthcare applications. These outcomes demonstrate that the proposed 1DCNN–BiLSTM approach is a reliable, interpretable, and practical tool for the early detection of cardiovascular risk in diabetic patients.

This package includes Python code for implementing the basic concepts and functions of the Numerical Atom Method proposed by Zhen (Leo) Liu. These six small examples implement the definitions and tran…

This package includes Python code for implementing the basic concepts and functions of the Numerical Atom Method proposed by Zhen (Leo) Liu. These six small examples implement the definitions and transfer convention in "Numerical Atom Method Specification, Version 1.0.0". The specification and this software have separate Zenodo records, connected through their Related works entries. Cite the version-specific DOI on each record when using the corresponding definitions or code. No account or network connection is needed to execute the examples after installing the dependencies.

Ушбу мақолада доғларда манийнинг ҳосил бўлган муддатига қараб текширишда  24 соат, 2, 3, 4, 5, 6 ва 7 суткадан сўнг, кейинчалик эса 2, 3, 4, 5 ва 6 ҳафтадан кейин дока дока бўлагидаги намуналар с…

Ушбу мақолада доғларда манийнинг ҳосил бўлган муддатига қараб текширишда  24 соат, 2, 3, 4, 5, 6 ва 7 суткадан сўнг, кейинчалик эса 2, 3, 4, 5 ва 6 ҳафтадан кейин дока дока бўлагидаги намуналар суд биологик текршилган. морфологиясини баҳолаш учун1-4сутка ,5-6 ҳафта, 2-3 ойлар кесимида Шорр усулида сперматозоидларнинг турли сақланиш муддатларида бўялишини ўрганилган.

As more of the judicial documents are digitized and the information becomes more complex, it has become more difficult to provide accurate information retrieval and summarization in legal documents. W…

As more of the judicial documents are digitized and the information becomes more complex, it has become more difficult to provide accurate information retrieval and summarization in legal documents. While current keyword-based retrieval systems often miss out on semantically similar, but lexically distinct, precedents, generic neural summarization models may struggle to capture legal reasoning, statutory context, and long-range context. The restrictions are especially relevant in the Indian context where judgment writing is lengthy, structurally varied, precedents dependent and so is the influence of the specific terminology in these judgments. This paper presents the Legal Aware BigBird Longformer Encoder–Decoder (LawBird-LED), a transformer-based hybrid model designed for intelligent legal retrieval and automated legal summarization to tackle these challenges. The proposed system involves using semantic ontology-based search to retrieve legally relevant information and case-based reasoning (CBR) retrieval to retrieve the precedent aware contextual knowledge for improving legal understanding. Furthermore, the structure applies both sparse attention (BigBird) and contextual attention (Longformer) for long legal documents and preserves the global relationships across the document and the local semantic relationships within the document. The summarization approach is presented as a multi-layer summarized approach to gradually improve the summarization process, and a set of interpretability techniques for the retrieval and summarization decisions is presented for transparency. Experimental comparison with the legal summarization and transformer-based benchmark models shows the proposed framework is able to generate legal summaries containing meaning and content well, and with good context capture. The proposed LawBird-LED achieved a ROUGE-F1 score of 99.0%, BLEU score of 98.8%, Legal-Sim score of 99.1%, METEOR score of 98.9% and Coverage score of 99.2%, which showed good lexical alignment, semantic consistency, and the retention of legal concepts. An effective and interpretable legal information retrieval and automatic legal text summarization system based on semantic retrieval, precedent-guided reasoning and hybrid contextual attention. The results reveal that augmenting symbolic legal knowledge with precedent-based reasoning and long-context neural representation can facilitate the semantic retrieval and legal-summary fidelity. The study therefore shows how LawBird-LED can be an interpretable decision support system for access to legal information, as well as pointing out issues of cross-jurisdictional generalization, computational efficiency and expert validation that need to be further explored.

Aggregate result tables underlying the technical report Constructing newcomer cohorts from Wikimedia logs. Eight tables covering four Wikipedia language editions (Polish, Czech, Hungarian, Danish) for…

Aggregate result tables underlying the technical report Constructing newcomer cohorts from Wikimedia logs. Eight tables covering four Wikipedia language editions (Polish, Czech, Hungarian, Danish) for accounts registered in 2024: account creation types, edit-count distributions with and without the local-registration control, first-edit revert against abandonment after a single edit, odds ratios with and without an individual six-month observation window, revert tag identifiers, and the query index with Quarry permalinks. Extracted from the public Wikimedia database replicas via Quarry on 4 October 2026. No individual-level data. Column descriptions and provenance are given in the README.

Fon krizi yalnızca birkaç fonun ya da gevşek denetimin hikâyesi değil. Fonlar milyonların sıradan tasarruf aracına dönüştü; ama servet yayılmadı. Borç tabana yayılırken fon serveti tepede yoğunlaştı. …

Fon krizi yalnızca birkaç fonun ya da gevşek denetimin hikâyesi değil. Fonlar milyonların sıradan tasarruf aracına dönüştü; ama servet yayılmadı. Borç tabana yayılırken fon serveti tepede yoğunlaştı. Tasarrufun yönetimi aracılara devredilirken kayıp hanehalkı bilançosuna dönüyor.