Our Work

This repository contains the analysis code and author-generated numerical data underlying the results reported in the manuscript “Trigonometric Cyber-Biomedical Phase Embedding for IoMT Attack D…

This repository contains the analysis code and author-generated numerical data underlying the results reported in the manuscript “Trigonometric Cyber-Biomedical Phase Embedding for IoMT Attack Detection.” The deposited materials support the repeated stratified split analyses, model-performance evaluation, paired representation comparisons, phase-layer ablation analysis, attack-category evaluation, and phase-feature utilisation analyses reported in the manuscript. The original WUSTL-EHMS-2020 dataset is not redistributed in this repository and is publicly available from Washington University in St. Louis. Instructions for obtaining the source dataset and reproducing the analyses are provided in the included README file.

The paper presents Lythri, a family of on-device language models specifically designed for emotional companionship. The Lythri family consists of two models: Lythri 7B-A4B (7.46B total, 4.5B active), …

The paper presents Lythri, a family of on-device language models specifically designed for emotional companionship. The Lythri family consists of two models: Lythri 7B-A4B (7.46B total, 4.5B active), Lythri 4B-A2B (4.63B total, 2.3B active). For training stage, We utilize a three-stage training pipeline including supervised Continued Pre-Training (CPT) , fine-tuning(SFT) with self-adaptive thinking(when there's no system prompt, trigger thinking) and group relative policy optimization(GRPO) using a Gemma-2 27B reward judge. Our training dataset for chat collect by multi-turn dialogues with a new 5-state conversation simulation system. Our research also proposes a new quantization called Q4_0_8, which enables a 4-bit GGUF file enhanced by dot product acceleration, delivering Q6_K-level performance with 1.16-1.76× faster prefill in most of tested cases. 

This dataset contains supplementary material for our work on CGIs in the vertebrate genome.

 Investment projects routinely miss their planned cost, schedule and benefit targets, yet risk identification—the first and most consequential stage of risk management—remains weakly …

 Investment projects routinely miss their planned cost, schedule and benefit targets, yet risk identification—the first and most consequential stage of risk management—remains weakly grounded in empirical evidence. This study examines how risks materialise in real-world investment projects and which failures of risk identification best explain the observed outcomes. Using a structured secondary-data analysis, we synthesise ex post evidence from six empirical studies: a database of more than 16,000 completed projects, three sector datasets (transport, dams, electricity infrastructure) and a survey of 701 project managers. Outcomes are compared by prevalence, magnitude, dispersion and, where reported, schedule and benefit indicators; coefficients of variation and mean-to-median ratios are derived from published statistics. The results show that overruns are the norm (three-quarters to nine-tenths of projects in the sector datasets exceed their cost estimates; only 8.5% of projects meet both cost and schedule targets), that outcome distributions are strongly right-skewed (mean cost overrun of 96% against a median of 27% for large dams), that forecasting accuracy has not improved over time, and that even moderate risk-management planning weakens the negative effect of risk on project success. We propose a three-layer risk identification gap framework (epistemic, behavioural and structural), derive testable propositions and formulate practical recommendations for investors and project sponsors. The findings are limited by reliance on published data dominated by large public and energy projects.

Raqamli fayl yaxlitligini baholashda usul tanlash tekshiriladigan savol va mavjud etalonga bog'liq. Tezisda xesh, elektron imzo, metama'lumot, format tuzilmasi va tasvir tahlili vazifa qamrovi bo'yich…

Raqamli fayl yaxlitligini baholashda usul tanlash tekshiriladigan savol va mavjud etalonga bog'liq. Tezisda xesh, elektron imzo, metama'lumot, format tuzilmasi va tasvir tahlili vazifa qamrovi bo'yicha solishtiriladi. Har bir usul uchun qo'llanish sharti va dalil chegarasi belgilanib, eng kam xarajatli yetarli tekshiruvni tanlash modeli taklif etiladi. Qamrov matritsasi baytlar mosligi, tasdiqlovchi kalit va o'zgarish sohasi bo'yicha savollarni birlashtirish uchun qaysi vositalar zarurligini ko'rsatadi. Etalon mavjud bo'lmagan holatning alohida ko'rib chiqilishi hamda yagona eksportdan hosil bo'lgan belgilarni birgalikda talqin qilish natijaning asosliligini ta'minlaydi.

This is a verbatim mirror of the official training set of the 2nd China Physiological Signal Challenge (CPSC 2019), "Challenging QRS Detection and Heart Rate Estimation from Single-Lead ECG …

This is a verbatim mirror of the official training set of the 2nd China Physiological Signal Challenge (CPSC 2019), "Challenging QRS Detection and Heart Rate Estimation from Single-Lead ECG Recordings". Contents (4,000 MATLAB files, 58 MB): data/data_00001.mat ... data_02000.mat : 2,000 single-lead ECG recordings from patients with cardiovascular disease; each recording is 10 seconds long, sampled at 500 Hz. ref/R_00001.mat ... R_02000.mat : the official R-peak (QRS) reference annotations for the 2,000 training recordings. Why this mirror exists: the official challenge website (http://2019.icbeb.org/Challenge.html) is no longer reachable, which made the training set difficult to obtain for reproducible research and for continuous integration. This record is hosted so that downstream software (e.g. the torch_ecg framework, https://github.com/DeepPSP/torch_ecg) and its test pipeline can fetch the data from a stable, unrestricted source. The file is redistributed unmodified. Integrity:MD5        6bf346bde7962f28f7464dca9c35a252SHA-256 1a838bcab24d5b93de5dfebd0c9f5b43b9a621415a6472f2056f6b79c7d5c7fc License: The underlying data were produced by the CPSC 2019 challenge organizers; no additional rights are granted by this mirror. Please refer to the original challenge terms before use. Citation: Please cite the original challenge (official website above; see also DOI 10.1166/jmihi.2019.2800) and this record.

Mazkur tezisda bank kapitalining yetarlilik darajasini baholash va uni boshqarish strategiyalari tahlil qilingan. Bank kapitalining moliyaviy barqarorlikni ta'minlashdagi o'rni, Bazel prudensial talab…

Mazkur tezisda bank kapitalining yetarlilik darajasini baholash va uni boshqarish strategiyalari tahlil qilingan. Bank kapitalining moliyaviy barqarorlikni ta'minlashdagi o'rni, Bazel prudensial talablari, kapital tarkibi va unga ta'sir etuvchi asosiy omillar yoritilgan. Shuningdek, O'zbekiston tijorat banklari misolida kapital tuzilmasi va uning dinamikasi baholanib, kapital boshqaruvini takomillashtirish yo'nalishlari ko'rsatib berilgan.

Timely identification of suitable blood donors is a critical operational challenge because donor availability, blood-group compatibility, geographical distance, and response behavior can change rapidl…

Timely identification of suitable blood donors is a critical operational challenge because donor availability, blood-group compatibility, geographical distance, and response behavior can change rapidly. This paper presents BloodMap AI, a web-based intelligent blood donation management framework that combines compatibility-aware donor ranking, location-based scoring, blood-demand forecasting, donor eligibility rules, persistent data management, and an extensible machine-learning/MLOps service. The frontend is implemented with React, Vite, TypeScript, and Supabase, while a Python FastAPI service provides protected matching, forecasting, retraining, health, and Prometheus monitoring endpoints. The donor matching component first removes incompatible donors and then considers blood-group compatibility, Haversine distance, and current availability; an XGBoost-based ranking model can be loaded when a trained model artifact is present, with a deterministic rules-based fallback for service resilience. The demand-forecasting interface aggregates actual blood-request records and produces a short-horizon forecast with lower and upper bounds. The repository also includes data extraction, preprocessing, model-training, model registration, drift monitoring, automated tests, and GitHub Actions workflows for continuous integration and weekly retraining. Functional validation of the uploaded project passed six unit tests covering compatibility, donor ranking, and forecasting behavior. The work demonstrates a practical architecture for AI-assisted blood-resource coordination while identifying the need for larger real-world datasets and statistically validated model benchmarks.

This dataset accompanies the manuscript “Guild: A flexible end-to-end framework for protein–ligand binding prediction.” It contains the data used to perform large-scale (bulk-mode) p…

This dataset accompanies the manuscript “Guild: A flexible end-to-end framework for protein–ligand binding prediction.” It contains the data used to perform large-scale (bulk-mode) protein–ligand binding predictions with Boltz2, DiffDock, and AutoDock Vina. The primary use case focuses on G protein–coupled receptors (GPCRs), including curated combinations of receptor structures with known ligands sourced from the Protein Data Bank (PDB), as well as decoys and experimentally validated binders defined in the manuscript. In addition, the dataset incorporates two broader chemical classes—natural products and synthetic compounds—to support extended benchmarking and generalization analyses.   Summary of the files present below: case_study/ – GPCR case study full_combinations_table.csv: the 270,000 protein–ligand combinations of the case study. Each of the 133 GPCR targets is paired with 1,000 natural products and 1,000 synthetic molecules, and each row gives the receptor, the ligand SMILES and the ligand class. It replaces the file of the same name in the previous version, which listed the wrong natural-product sample. CORRECTION.md: what was corrected and how the correction was checked. figure3_normalization/ – score normalization comparison vinarun_scores.txt: AutoDock Vina scores, ranks and rank percentiles for every case-study combination, including the 1,000 decoys per target that serve as the ranking reference. knownbinders_scores.txt: AutoDock Vina scores for 655 known binders (pChEMBL > 6) on 131 of the targets, each ranked against that target's decoys. benchmark_three_targets/ – per-method benchmark guild_scores.txt: 165 protein–ligand pairs, made of 5 known binders and 50 property-matched decoys for each of CB1 (7V3Z), SMO (6OT0) and EP3 (8GDC). Each pair has scores from AutoDock Vina, GNINA, KarmaDock, DiffDock and Boltz-2, the Vina and GNINA rescores of DiffDock and Boltz-2 poses, PoseBusters validity, per-score rank percentiles, and Guild's combined score (global_rp_score). native_ligand_rmsd.tsv: RMSD of each method's pose of the co-crystallized ligand to the crystal pose. batch_progress.log: timestamps of the benchmark run, which give the runtimes. Convention: the rp_* and global_rp_score columns store rank / N, so lower is better. The manuscript reports 1 − rank / N, so that higher is better.   pdbs.zip - PDB files for the GPCR proteins used in the case study gpcrdb_protein_data.csv - Curated metadata on GPCR proteins in the case study

Version 2, correction of record. The first version cited M. Land, Pair production in classical Stueckelberg-Horwitz-Piron electrodynamics, as J. Phys. Conf. Ser. 845, 012025 (2017). That journal refer…

Version 2, correction of record. The first version cited M. Land, Pair production in classical Stueckelberg-Horwitz-Piron electrodynamics, as J. Phys. Conf. Ser. 845, 012025 (2017). That journal reference belongs to a different paper by the same author. The correct reference is J. Phys. Conf. Ser. 615, 012007 (2015), doi:10.1088/1742-6596/615/1/012007. The work replicated in this paper was transcribed from the preprint arXiv:1604.01625, which is that paper, so no result, figure, or number changes. Only the printed journal reference was wrong.Can a local hidden dynamics observed through a time fold produce pair-creation statistics with the Schwinger exponential's field dependence, without inserting it? For one declared family the sealed answer is no. In a Sauter-slab energy-transfer model, time-reversal suppression is organized by a Gaussian measure tail in a probe-measured effective gap, and the Schwinger-shaped alternative fails to compete on held-out data. The paper reports the complete chain: a four-iteration pilot whose every failure became a design constraint (including a critical-field manifold separating deterministic from fluctuation-driven reversal), a first preregistered run whose neither-axis verdict exposed the designer's own bars as miscalibrated, a correction rule declared in writing before any second attempt, and a second preregistered run passing both corrected bars with room (generalization ratio 0.38 against a bar of 2, competing axis 3.14 times worse against a bar of 2, prescription controls exact per-member). The verdict binds one family only; the open question, whether any constructible hidden-dynamics family has a non-measure-tail exponent, is stated with what a claimed positive would owe.Fifth paper of the observation-theory campaigns; companion to 10.5281/zenodo.21790096 (the Cayley-pole replication whose finding became this paper's closure clause) and 10.5281/zenodo.21789011. All numbers trace to append-only hashed evidence records under a six-seal ledger.