Projects

### Overview: The IAB Content Taxonomy Mapper is a tool designed to help you migrate from IAB Content Taxonomy 2.x to 3.0 by mapping your existing 2.x codes to the new 3.1 format. This process includes deterministic matching, fuzzy matching, and optional semantic enhancements, ensuring compatibility with OpenRTB and VAST standards. ### Key Features: 1. **Input:** Your existing 2.x codes/labels (CSV or JSON). 2. **Process:** A deterministic → fuzzy → optional semantic pipeline. 3. **Output:** Valid IAB 3.1 IDs ready for OpenRTB/VAST integration, with optional vector attributes and SCD awareness. ### Tech Stack: - Python ### Why It Matters: - Perfect for ad tech teams, content platforms, and anyone migrating to IAB 3.0. - Ensures compatibility with OpenRTB and VAST standards, making it easier to integrate with modern ad tech platforms. ### Repository: - **Repository:** [IABTechLab/iab-mapper](https://github.com/IABTechLab/iab-mapper) - **GitHub Stars:** 23 - **Primary Language:** Python - **GitHub Description:** Local IAB Content Taxonomy Mapper (2.x → 3.0). Deterministic → fuzzy → optional embeddings. Exports OpenRTB & VAST-ready category IDs with vector attributes and SCD awareness. ### Example: ```bash # Input: 2.x codes "1-4","Sports" "2-12","

### Overview: msiFlow is a collection of mass spectrometry imaging workflows designed for automated and reproducible analysis of multimodal mass spectrometry imaging and microscopy data. The latest version, v2.0.0, includes improvements in mass recalibration, lipid annotation, batch correction, quality control visualizations, and computational efficiency. ### Key features: - **Mass Recalibration**: Enhances the accuracy of mass measurements. - **Lipid Annotation**: Adds lipid information to the analysis for better context. - **Batch Correction**: Corrects for systematic errors across multiple samples. - **Quality Control Visualizations**: Provides visual insights into the quality of processed data. - **Computational Efficiency**: Optimizes the processing pipeline for faster analysis. ### Tech stack: Python ### Why it matters: msiFlow offers a robust and scalable solution for mass spectrometry imaging data analysis, making it essential for researchers in the field of proteomics and metabolomics. The improvements in workflow efficiency and accuracy contribute to more reliable and reproducible results, which is crucial for scientific research and clinical applications.

### Overview: ETCHR (Editing To Clarify and Harness Reasoning) is a question-conditioned, reasoning-aware image editor designed to serve as a decoupled visual reasoning assistant for Multimodal Large Language Models (MLLMs). By separating the specialized image editor from the understanding model, ETCHR addresses the critical bottleneck where a purely textual chain of thought fails in fine-grained focus or complex spatial transformations. ### Key features: - **Decoupled & Plug-and-Play:** ETCHR functions as a separate module, allowing it to assist diverse downstream MLLMs without requiring any task-specific fine-tuning on the understanding models themselves. - **Naturally Reflective Pipeline:** Introduces an Edit-Verify-Reason inference mechanism where the understanding model filters out noisy or flawed edits, reverting safely to the original image when verification fails. ### Tech stack: Python ### Why it matters: ETCHR bridges the gap between purely textual chains of thought and visual reasoning, making it a valuable tool for enhancing the capabilities of Multimodal Large Language Models (MLLMs) in various downstream applications, such as fine-grained perception, chart understanding, logic reasoning, jigsaw restoration, and 3D understanding.

Overview: Automatically describe images sent by users on popular media platforms, incredibly useful for the visually impaired and for complicated imagery. Key features: - Image description for Discord, Slack, Messenger, Telegram, and other platforms. - Integration with OpenAI's GPT-4-Vision for accurate image interpretation. - Support for setting up a bot and a demo server. - Easy setup using Python 3.9 and above, with environment variables and Discord token. Tech stack: Python Why it matters: Described is a valuable tool for visually impaired users and those dealing with complex imagery, offering a comprehensive understanding of images without the need for visual perception.

### Overview: This project provides benchmarks and notes for running modern language models (LLMs) on NVIDIA's Tesla V100 GPUs, specifically focusing on the 8x V100-32GB configuration with CUDA 12.6. The primary goal is to explore the capabilities and performance of these older GPUs with the latest version of vLLM, which supports modern model families and 7x V100-32GB setups. ### Key Features: - **Modern Model Support**: Benchmarks for 7 different model families, including Qwen3.5, Qwen3.6, and GLM models. - **GPU Configuration**: 8x NVIDIA V100-32GB SXM2 (sm_70) with CUDA 12.6. - **Engine Versions**: vLLM 0.19.x and 0.21.x, both source-built on CUDA 12.6. - **FP8 + MoE Plugin**: A custom FP8 + MoE plugin for the V100-32GB configuration. - **FlashAttention-V100**: Engine-specific ABI builds for vLLM 0.21 and 0.19. - **Baseline Matrices**: Full per-engine matrices for 0.19 and 0.21 engines. - **FP16 MoE Bug**: A fix for the 4-9x slowdown in MoE on V100 from a default setting. - **FP8 Plugin Performance**: Custom

Overview: This project involves a comprehensive **Natural Language Processing (NLP)** analysis of Persian social media sentiment over a 5-year historical period, from 2020 to 2025 AD, using Telegram. The primary objective is to analyze public mood trends on Telegram, utilizing Large Language Models via Ollama, and classify nearly 1 million posts into precise emotional categories, handling linguistic nuances of the Persian language. Key features: - Historical data mining from 5 major public channels - Persian NLP pipeline with text normalization using Hazm - Custom-engineered Persian System Prompt with Gemma 3 for distinguishing deep cultural emotions - Visual analytics including time-series trends and "Hope vs. Despair" statistical ratios Tech stack: Python, Google Gemma 3 (27B), Ollama Why it matters: This project provides valuable insights into public mood trends, cultural nuances, and sarcasm in Persian social media, offering a deep dive into the emotional landscape of a major social media platform.

### Overview: MiloQ/MELD-Sentiment-Analysis 是一个多模态情感分析项目，使用 PyTorch 框架结合文本和语音模态信息进行情感分类。该项目使用了 MELD 数据集，包含《老友记》的对话文本和其对应的语音和视频。目前项目重点考虑文本和语音两种模态信息的多模态特征融合。 ### Key features: - MultiModal Sentiment Analysis (Text and Audio) - Pytorch 框架 - Huggingface 的 Transformers 库 - 文本特征抽取：使用 BERT 模型 - 语音特征抽取：使用 HuBERT 模型 - 特征融合：文本和语音特征的 concat 和 cross modal attention ### Tech stack: PyTorch ### Why it matters: 本项目通过结合文本和语音模态信息，为情感分析提供更全面和准确的视角，特别是在处理多模态数据集时。通过多模态特征融合，提高了模型的分类效果，特别是在识别情绪方面。

Overview: ArtifyX is an AI-powered art generator that allows users to generate up to five images from a single prompt in their terminal. Key features: - Generate five AI-generated images from one prompt - Supports modern OpenAI SDK v1+ - Uses DALL·E model choice - Saves up to 10 images as timestamped PNG files - Works outside Termux with live config validation - Supports every major Linux family - Verifies dependencies before claiming success Tech stack: Python Why it matters: ArtifyX simplifies the process of generating art by allowing users to describe their desired image in the terminal, making it accessible and efficient.

### Overview: Turbopuffer GUI is a desktop application designed for developers to interact with the Turbopuffer vector database. It provides a user-friendly interface for managing connections, browsing namespaces, exploring documents, and configuring advanced features like aggregation queries and schema design. ### Key Features: - **Connection Management**: Securely manage multiple Turbopuffer connections with encrypted API key storage. - **Namespace Browser**: View and manage all namespaces in your Turbopuffer instance. - **Document Explorer**: Browse, search, and filter documents with an intuitive interface. - **Schema Designer**: Visual schema design tool for configuring vector dimensions, full-text search, and attribute indexes. - **Advanced Filtering**: Build complex filters using a visual filter builder or raw query mode. - **Aggregations**: Run aggregation queries with group-by support. - **Dark Mode**: Terminal-inspired dark theme. ### Tech Stack: - **Electron**: Cross-platform desktop framework. - **React 19**: UI framework. - **TypeScript**: Type safety. - **Vite**: Build tool. - **Tailwind CSS**: Styling. - **Radix UI**: Accessible component primitives. - **Zustand**: State management. ### Why It Matters: Turbopuffer GUI is an essential tool for developers looking to integrate vector and full-text search capabilities into their applications. It simplifies the process of managing connections, browsing namespaces, and exploring documents, making it easier for developers to leverage the power of Turbopuffer in their projects.

Overview: Formation MariaDB 12.3 LTS est un guide progressif dédié aux débutants et experts, couvrant divers aspects de MariaDB, incluant SQL, index, transactions, réplication, Galera, MaxScale, DevOps/Kubernetes, et MariaDB Vector. Key features: - 20 modules progressifs - MariaDB 12.3 LTS version de référence - MariaDB Vector IA/RAG intégration - Production-ready production, monitoring, HA, architectures modernes - Pratique, avec 200+ exemples SQL, configurations, checklists - 9 annexes, glossaire, commandes, références techniques - En français, accessible pour tous niveaux Tech stack: PHP, Markdown, HTML, CSS, JavaScript, Node.js, Express.js, MongoDB, MySQL, Docker, Kubernetes, Git, Docker Compose, Helm, Ansible, Terraform, Terraform Cloud, AWS, Azure, GCP, Terraform Registry, Terraform Registry Provider, Terraform Registry Provider Provider, Terraform Registry Provider Provider Provider, Terraform Registry Provider Provider Provider Provider, Terraform Registry Provider Provider Provider Provider Provider, Terraform Registry Provider Provider Provider Provider Provider Provider, Terraform Registry Provider Provider Provider Provider Provider Provider Provider, Terraform Registry Provider Provider Provider Provider Provider Provider Provider Provider, Terraform Registry Provider Provider Provider Provider Provider Provider Provider Provider Provider, Terraform Registry Provider Provider Provider Provider Provider Provider Provider Provider Provider Provider, Terraform Registry Provider Provider Provider Provider Provider Provider Provider Provider Provider

Overview: SWE-Serve is an agentic benchmark designed to evaluate Large Language Models (LLMs) and agents on production inference engineering tasks, focusing on correctness and real-world scenarios. Key features: - 53 production inference-engineering tasks derived from merged SGLang pull requests - 41 GPU tasks and 12 CPU tasks - Covers integration into an inference system, persistence, and concurrent requests - Runs on Linux x86_64 with Docker and Python 3.12+ - Requires a machine with NVIDIA H100 80GB SXM5, at least 20 CPU cores, and 128 GiB of RAM - Supports closed-book runs for open-book benchmark results Tech stack: Python Why it matters: SWE-Serve provides a comprehensive benchmark for evaluating LLMs and agents in a production environment, ensuring their performance aligns with real-world use cases. Repository: NVIDIA/swe-serve GitHub stars: 23 GitHub description: SWE-Serve: an agentic benchmark of 53 production inference-engineering tasks derived from merged SGLang pull requests, run with Harbor.

Overview: PerCoV2 is an ultra-low bit-rate perceptual image compression system that leverages a novel 1D query-based tokenizer and dedicated entropy model to achieve state-of-the-art compression performance across various benchmarks. Key features: - Novel 1D query-based tokenizer - Dedicated 1D masked entropy model - State-of-the-art compression performance on MSCOCO-30k and general-purpose datasets Tech stack: PyTorch Why it matters: PerCoV2 offers a comprehensive solution for ultra-low bit-rate perceptual image compression, addressing the challenges of maintaining realistic reconstructions at low bit-rates while avoiding structured artifacts.

### Overview: This repository is an implementation of a GPT-style Large Language Model built from scratch using PyTorch. It follows the concepts and techniques described in Sebastian Raschka's "Build a LLM from Scratch" book, focusing on fundamental aspects of building, training, and fine-tuning language models. ### Key features: - **Data Preprocessing & Tokenization**: Custom tokenizers for text conversion to numerical tokens. - **Attention Mechanisms**: Basic self-attention and causal attention for generating text. - **Model Architecture**: Full implementation of GPT with configurable layers and efficient multi-head attention. - **Training Pipeline**: Complete training loop with evaluation and loss calculation utilities. ### Tech stack: Python ### Why it matters: This project provides a comprehensive guide to building and fine-tuning language models from scratch, making it valuable for developers looking to understand and implement advanced natural language processing techniques.

Overview: The KoCC3M dataset is a Korean translation of Google's Conceptual Captions dataset, focusing on creating machine-translated captions for vision-image datasets. Key features: - Machine-translated captions for images - Gold-label translation captions for human supervision Tech stack: Python, PyTorch, Hugging Face Transformers Why it matters: This dataset aims to further enhance the creation of human-supervised gold label translation captions, expanding the scope of image captioning research.

Overview: DeraineDB is a high-performance, hardware-first vector database engine built in Zig and Go. It offers sub-millisecond HNSW search with 64-bit metadata filtering and a 1.8MB binary footprint, making it ideal for local AI applications and retrieval-augmented generation. Key features: - Sub-millisecond HNSW search - 64-bit metadata filtering - 1.8MB binary footprint - Hardware-first storage layer Tech stack: Zig and Go Why it matters: DeraineDB is designed to run on bare-metal, achieving speeds comparable to traditional databases while maintaining a minimalistic design, making it suitable for edge computing and local AI applications.

### Overview: HatReD is a public repository containing a dataset and code implementation for the paper "Decoding the Underlying Meaning of Multimodal Hateful Memes" (IJCAI'23). This dataset and code are based on the original Facebook AI's Hateful Memes dataset and aim to decode the underlying meaning of multimodal hateful memes. ### Key Features: - HatReD dataset with annotations for social targets and reasoning. - Code implementation for various encoder-decoder models. - Baseline experiments and training scripts for different models. - Evaluation scripts for model performance on the HatReD dataset. ### Tech Stack: Python ### Why It Matters: HatReD provides a comprehensive resource for researchers and developers interested in analyzing and understanding the underlying meanings of multimodal hateful memes. The dataset and accompanying code offer a valuable tool for academic research and practical applications in natural language processing and sentiment analysis. ### Repository: Social-AI-Studio/HatReD GitHub Stars: 23 ### Primary Language: Python ### GitHub Description: The repository contains HatReD dataset and code implementation for the paper "Decoding the Underlying Meaning of Multimodal Hateful Memes" (IJCAI'23).

Overview: TURN is an entropy-based algorithm designed to automatically optimize the temperature parameter in multi-sample inference strategies like Majority Voting and Best-of-N. This approach aims to enhance the performance of language models without the need for labeled validation sets. Key features: - Automatically predicts the optimal temperature for given aggregation strategies. - High correlation between predicted and grid-search temperatures. - Purely entropy-driven, removing reliance on labeled validation sets. Tech stack: Python Why it matters: TURN provides a practical solution for optimizing temperature in language models, potentially improving their performance without the need for extensive labeled data. Repository: StigLidu/TURN Primary language: Python GitHub stars: 23 GitHub description: [ICML2025] Official Repo for Paper "Optimizing Temperature for Language Models with Multi-Sample Inference"

Overview: Kitty is a plug-and-play KV-cache system designed for HuggingFace Transformers, focusing on accurate and efficient 2-bit KV-cache quantization. This innovative system is inspired by channel-wise patterns and quantization sensitivity, offering a dynamic channel-wise precision boost for improved performance. Key features: - Accurate and efficient 2-bit KV-cache quantization - Dynamic channel-wise precision boost for improved accuracy and efficiency Tech stack: Python Why it matters: Kitty addresses the challenges in quantization for large language models by providing a practical and effective solution, ensuring high accuracy and low resource consumption, making it a valuable tool for developers and researchers working with HuggingFace Transformers.

### Overview: SynapseKit is a minimal, async-first Python framework designed specifically for production LLM applications. It provides a straightforward and efficient way to integrate LangChain with various LLM providers, vector stores, data loaders, and agent tools, all while ensuring no hidden dependencies or global state. ### Key Features: - **Minimal Dependencies**: SynapseKit relies on only two hard dependencies, ensuring a clean and lightweight setup. - **Async-Native RAG, Agents, and Graph Workflows**: Utilizes async/await for seamless integration with LangChain and other asynchronous APIs. - **No Magic, No SaaS, No Bloat**: The framework is designed to be simple and straightforward, without any hidden chains or global state. ### Tech Stack: - **Python**: The primary language used for development and implementation. ### Why It Matters: SynapseKit simplifies the integration of LangChain with various LLM providers, vector stores, data loaders, and agent tools, making it easier for developers to build production-grade LLM applications without the complexity of magic or unnecessary dependencies.

### Overview: HandWrite - 一个基于 C++/Qt6 的手写体文本生成器。它将普通文本转换为逼真的手写图像，支持多页、并行渲染、字符级属性覆盖和丰富的参数调节。完美生成手写作业、笔记、信件等个性化文档。 ### Key features: - **纸张纹理**：支持 6 种纸张模板：横线纸、方格纸、田字格、作文纸、点阵纸、纯色背景，可调透明度。 - **背景图片 + 透视校准**：自定义背景，点击纸的四个角自动校准文字到歪斜纸张上。 - **笔触效果**：模拟笔画粗细扰动、墨水洇染、涂改模拟。 - **字体**：手写字体库、混合字体、字符级别覆盖。 - **排版增强**：段首缩进、段间距、中文标点保留/转 ASCII、竖排、文字变形。 - **复现**：随机种子、可拖拽分隔条、多预设管理、长任务进度对话框可取消。 - **构建**：依赖 MSYS2、GCC、CMake、Qt 6.11，支持多平台构建。 ### Tech stack: Qt 6.11, C++ 17, CMake 3.16+, GCC 16+ ### Why it matters: HandWrite 是一个强大的工具，适用于生成各种手写体