QUELL — reproducibility package (v4.0). This repository accompanies the paper “QUELL: An On-Device Benchmark of Quantized Edge LLMs versus Classical Machine Learning for IoT Intrusion Detection.” QUELL fine-tunes decoder-only LLMs (GPT-2-medium, Qwen2.5-0.5B/1.5B, Llama-3.2-1B) on network-flow records serialized as text, quantizes the Qwen2.5-1.5B detector, deploys it on an NVIDIA Jetson Orin Nano Super (GGUF/llama.cpp), and compares it with random-forest and XGBoost detectors on Edge-IIoTset, CICIoT2023 and N-BaIoT under a leakage-controlled, multi-seed protocol. Axes: detection quality, on-device latency/energy/memory, quantization retentio
Şüyun, Süleyman Burçin, Aktürk, Fatih
What's New in v2.0.0 Chanakya - chat agent [x] On-Demand Deep Biological Evaluations : Query PubMed articles and generate structured scientific syntheses covering executive findings, assay workflows, molecular/epigenetic mechanisms, and facility takeaways. [x] Dual LLM Provider Support : Compatible with both Groq (ultra-fast Llama/Mixtral/GPT-OSS models) and Google Gemini . Pure Client-Side & Zero-Exposure Privacy [x] Direct User API Keys : Users securely provide their own API keys, stored solely within their browser's localStorage . [x] Zero Server Overhead & Zero Secret Leaks : No server-side proxies, no stored user logs, and no shared API
Das, Jyotirmoy
OGX (Open GenAI Stack), formerly Llama Stack, is an open-source, vendor-neutral generative AI application server and Python library with interchangeable backend providers. It supports retrieval-augmented generation, conversational agents, and tool-calling workflows, allowing applications to change inference engines and vector databases through configuration rather than application rewrites. OGX provides server-side Responses API orchestration, persistent conversations, vector stores, and configurable safety and authorization controls. This archive contains the source code for OGX v1.5.0, released on October 9, 2026, corresponding to Git commi
Arceo, Francisco Javier, Han, Sébastien, Farrellee, Matthew, Doern, Charlie, Tang, Yuan, Higgins, Derek, Narsing, Varsha Prasad, Sim, Gordon, Kamenani, Sumanth, Browning, Ben, Murthy, Raghotham
SiteFootprint estimates the operational carbon emissions and on-site water use of training an AI model at any location. It computes data-centre cooling hour by hour from real weather, combines it with annual grid carbon intensity, and reports results with 95% uncertainty ranges and paired comparisons between locations. The cooling model was calibrated on published Google campus PUE and tested on unseen campuses (mean absolute error 0.010), and the energy model reproduces Meta's disclosed Llama 3.1 405B training hours to within 5%. Available as a Python package, a command-line tool and a web app, with citable exports.
Oladeji, Lawrence
Code, derived results and figures for two papers on selection and geometry in machine learning. Paper 1, Selection Under an Imperfect Verifier, develops an extreme-value theory of test-time compute: best-of-n delivers a concomitant of an extreme order statistic, which yields the log-linear gain curve, a proof that long-tailed verifier error forces a finite optimal sample count, an unbiased U-statistic estimator of the whole gain curve from one pool, and exact greedy budget allocation. It is tested on 128,000 real completions (MATH-500, Llama-3.2-1B-Instruct, 8B process reward model). Paper 2, The Spectrum, Not the Norm, analyses orthogonalise
Singham, Shashvat
Code and container definitions for the AI application benchmarking framework presented in "AI Application Benchmarking: Power-Aware Performance Analysis for Vision and Language Models" ( https://arxiv.org/abs/2603.16164 ). The framework measures throughput and energy efficiency of representative computer vision workloads (ResNet-50, ViT-L/16, Stable Diffusion v2) and large language model workloads (LLaMA 3 8B pre-training with LitGPT and inference serving with SGLang) under systematic GPU power-cap sweeps. Includes Apptainer container definitions and configurations used to evaluate NVIDIA H100, NVIDIA H200, and AMD MI300X accelerators on a pe
Mayr, Martin, Wind, Sebastian, Lukas Schröder, Moradi, Mohammadmoein, Hager, Georg, Koestler, Harald, Wellein, Gerhard
Breaking Changes meta/llama-3.1-405b-instruct-turbo will no longer be used as a LLM-as-judge in Omni-MATH, WildBench, HarmBench, XSTest, and Anthropic Red Team scenarios due to deprecation by Together AI (#4233) Models Deprecate meta/llama-3.1-405b-instruct-turbo (#4233) Scenarios Release Arabic Legal scenario (#4226, #4236, #4237) Release Arabic Content Generation scenario (#4224, #4229, #4236, #4237) Migrate Unitxt integration to use new Unitxt API (#4234) Remove meta/llama-3.1-405b-instruct-turbo from annotators (#4233) Fix AutobencherSafetyScenario parsing the file path string as JSON (#4206) Contributors Thank you to the following contri
Tony Lee, Yifan Mai, Percy Liang, Dilara Soylu, Rishi Bommasani, Dimitris Tsipras, Yian Zhang, Deepak Narayanan, Josselin Somerville Roberts, Ryan Chi, Eric Zelikman, Farzaan Kaiyom, lxuechen, Frieda Rong, Brian W. Goldman, Drew Arad Hudson, Ananya Kumar, MiguelAFH, Yuhui Zhang, Ben Newman, Nathan Kim, Keshav Santhanam, Haoqin Tu, fladhak, Huaxiu Yao, Tianyi, Qian Huang, Michi Yasunaga
Comparing LLM based approaches for predicting Mg dissolution modulators. There is a Readme with brief instructions. Copyright (c) 2025 Busch et al. If you have any questions please contact us at , or . For further details refer to our publication ( https://doi.org/10.1016/j.corsci.2025.113080 ). Please cite this work if you use any of the code or ideas provided.
Busch, Matthias, Tacke, Marius, Lamaka, Sviatlana, Zheludkevich, Mikhail, Linka, Kevin, Cyron, Christian Johannes, Feiler, Christian, Aydin, Roland C.
CICIoT2023 üzerinde klasik makine öğrenmesi modelleri (Lojistik Regresyon, Rastgele Orman, XGBoost) ile sıfır-atış bir büyük dil modelinin (Gemini) doğruluk, genelleme, gecikme ve maliyet açısından karşılaştırması; çapraz-tür genelleme, çapraz-küme (NSL-KDD) doğrulaması, istem ablasyonu, ikinci LLM ailesi (Llama) ve SHAP açıklanabilirliği dâhil.
Küten, Ferhat, Özalp, Ahmet Nusret
Mushika is an offline python-based administrative utility designed for local system diagnostic logging and routine process monitoring. It provides basic system-level process auditing and network port tracking through a local SQLite database configuration. Core Features: Local Activity Log: Audits network port state and logs changes to an offline SQLite database file ( sutra_transitions.db ). Log Segmentation: Uses a dual-process layout to organize telemetry logs and user alerts into separate tracking files. Telemetry Indicators: Captures and writes routine system parameters (such as system coherence and entropy indicators) to local metrics lo
Quintana, Zachariah
EcaVault est une application web progressive (PWA) conçue spécifiquement pour les inspecteurs et cadres des Établissements Cantonaux d'Assurance (ECA) suisses. Elle intègre un pipeline d'intelligence artificielle multi-modal combinant l'analyse visuelle de photos via Gemini Flash (Google), la génération de rapports institutionnels via Groq Llama-3 , et un algorithme de scoring de risque déterministe (score 0–100) enrichi par LLM. L'architecture offline-first (IndexedDB + Background Sync API) garantit l'utilisation en zone sans réseau. Le système implémente un modèle de données spécifique aux ECA (numéro de police ECAB, valeur assurée CHF, niv
Stadelmann, Sébastien
Overview Antrocon is a multi-agent orchestration system for large language models (LLMs), designed for research into autonomous agent collaboration, structured dialog management, and long-term memory retrieval. It coordinates two autonomous LLM agents — a Builder (Agent A) and a Critic (Agent B) — in a turn-based dialog loop, while a human Operator (the "Papa Agent") monitors and steers the system through a real-time web dashboard. Architecture The system runs locally on a Windows host with two instances of llama.cpp serving quantised open-weight models via an OpenAI-compatible API. A central Python orchestrator manages dialog flow, tool exec
Massinger, Corbinian Elias
