Projects

Overview: ToolQA is a dataset designed to evaluate the capabilities of large language models (LLMs) in answering challenging questions with external tools. It includes eight real-life scenarios and two levels (easy/hard) across eight distinct domains. Key features: - Questions are selected to ensure LLMs have little chance of memorizing and answering correctly. - The majority of questions require compositional use of multiple tools. - Easy and Hard levels are distinguished based on the length of toolchains. - A thorough diagnosis and analysis of in-context tool-augmented LLMs are conducted in the paper. - Data from 8 domains, including temporal, spatial, social, scientific, and personal, with 55 templates and 800 questions in total. Tech stack: Jupyter Notebook Why it matters: ToolQA provides a comprehensive evaluation framework for LLMs, focusing on their ability to interact with external tools effectively, making it valuable for researchers and developers aiming to improve LLM performance in real-world scenarios.

Overview: AgentSwarms is a unified agentic AI and data platform that provides both AI agents and multi-agent swarms, as well as ETL pipelines, a lakehouse, BI dashboards, and no-code machine learning capabilities. It is built on self-hosted infrastructure, allowing users to control their model keys and data. Key features: - Multi-Agent Swarms - AI Agents - RAG Knowledge Bases - ETL Pipelines - Lakehouse - BI Dashboards - No-Code Machine Learning Tech stack: TypeScript Why it matters: AgentSwarms offers a comprehensive solution for building and managing AI agents and data platforms, providing a unified and self-hosted solution that integrates AI agents, RAG knowledge bases, ETL pipelines, a lakehouse, BI dashboards, and no-code machine learning capabilities. The platform's governance ensures that model rules, spend budgets, row filters, and column masks are enforced server-side, providing a consistent and reliable experience for users.

Overview: This repository contains the code for the CVPR 2024 paper "PIGEON: Predicting Image Geolocations". The paper was authored by Lukas Haas, Michal Skreta, Silas Alberti, and Chelsea Finn at Stanford University. Key features: - Preprocessing and data loading - Model training and evaluation - Auxiliary data download Tech stack: Python Why it matters: This project showcases the implementation of a novel method for predicting image geolocations, which is crucial for applications requiring accurate location-based image analysis.

Overview: RelevanceAI is a platform that provides AI agents and tools for building an AI workforce. Key features: - Agents for automating tasks - Tools for executing predefined workflows - Knowledge sets for storing and retrieving data - Tasks for managing and tracking work progress Tech stack: Python Why it matters: RelevanceAI offers a comprehensive solution for integrating AI agents, tools, and knowledge into an AI workforce, making it easier to automate processes and improve productivity.

Overview: Material Palette is an innovative method for extracting materials from a single real-world image, including albedo, normals, and roughness, and generating texture concepts for 3D rendering. Key features: - Extracts texture concepts from image and region mask sets - Allows generation at different resolutions - Proposes a decomposition step using a trained model Tech stack: Python, PyTorch, Diffusers, Peft, PyTorch Lightning Why it matters: This method provides a powerful tool for extracting and generating materials from images, enabling advanced 3D rendering and texture synthesis.

Overview: This repository contains a collection of high-performance large language model implementations written with JAX, focusing on minimal and performant models. Key features: - DeepSeek R1 - Llama 4 - Llama 3 - Qwen 3 - Kimi K2 - OpenAI GPT OSS - NVIDIA Nemotron 3 Nano - Gemma 4 Tech stack: Pure JAX Why it matters: These examples showcase the efficiency and performance of JAX in creating large language models, making them valuable for developers looking to optimize their models with minimal code and resources.

Overview: This project utilizes deep learning techniques, specifically the YOLO algorithm, to detect vehicles in images. Key features: - Fine-tuning of a pre-trained YOLO model on a custom dataset - Implementation of the YOLO algorithm for vehicle detection - Utilization of GitHub for project management and collaboration Tech stack: Python Why it matters: This project demonstrates the application of deep learning in vehicle detection, showcasing the integration of YOLO with custom datasets for improved accuracy and efficiency. Repository: MaryamBoneh/Vehicle-Detection Primary language: Python GitHub stars: 285

### Overview: This repository is a curated list of real Micro SaaS tools categorized by niche, perfect for inspiration, research, and startup ideas. ### Key Features: - **AI & Automation**: Tools for AI-driven automation and security, such as Agent Security, AI Cover Letter Generator, and AI for Google Slides. - **Audio & Video**: Tools for audio and video editing, including AI Image to Text and AI Wedding Toast. - **Business**: Tools for business management and marketing, like Foundy and SalesMagic. - **Career & HR**: Tools for career development and HR management, such as Lisapet.ai and Poshwatch. - **Customer Support**: Tools for customer support, including Marblism and Poison Pill. - **Design**: Tools for design and creativity, like Aigazine and Composed. - **Developer Tools**: Tools for software development, including AI for Google Slides and Palm. - **Education**: Tools for education and learning, like Agent Security and Lisapet.ai. - **Finance**: Tools for financial management, including TaxTip.ai and Velory. - **Health & Lifestyle**: Tools for health and lifestyle, such as Zolltarifnummer.com. - **Legal & Services**: Tools for legal and business services, including Patrons and Zippy Delivery. - **Marketing**: Tools for marketing and advertising, like Shopme Agent and Voice Command. - **Productivity**: Tools for productivity and automation, including SalesMagic and Zippy Delivery. - **Social**: Tools for social media and communication, like Marblism and Zolltarif

### Overview: Q-Bench is a benchmark designed to evaluate multi-modality language models (MLLMs) on low-level vision and visual quality assessment. It includes three realms: perception (A1), description (A2), and assessment (A3). The benchmark uses two datasets: LLVisionQA for perception and LLDescribe for description, and provides an abstract evaluation code for assessment using public datasets. ### Key Features: - **LLVisionQA**: A benchmark dataset for low-level vision perception. - **LLDescribe**: A dataset for visual description tasks. - **Abstract Evaluation Code**: Provides an evaluation framework for MLLMs on the assessment task using public datasets. ### Tech Stack: Jupyter Notebook ### Why It Matters: Q-Bench offers a comprehensive benchmark for evaluating the performance of multi-modality language models in low-level vision and visual quality assessment. By providing clear datasets and an evaluation framework, it helps researchers and developers understand the capabilities and limitations of these models in practical applications. Repository: Q-Future/Q-Bench GitHub Stars: 285

Overview: GPT-Synthesizer is an AI-driven software design and development tool designed to collaborate with developers to create complete software projects. It uses LangChain to interpret and synthesize prompts into executable code, focusing on capturing design details that are often overlooked in single-prompt specifications. Key features: - Collaborative software project design and development - AI-driven synthesis of design specifications into executable code - Streamlined process for novice and expert developers - Versatile design space exploration and component compilation - User-friendly command-line interface and Streamlit-based user interface - OpenAI's `gpt-3.5-turbo-16k` as the default LLM Tech stack: Python, LangChain, OpenAI's `gpt-3.5-turbo-16k` Why it matters: GPT-Synthesizer simplifies the complex task of software project design and development by focusing on capturing design details that are often overlooked in single-prompt specifications. This makes it an invaluable tool for both novice and expert developers, ensuring a more efficient and effective development process.

### Install on Mac (option 2) ``` # Install Homebrew if you don't have it already /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)" # Install Elasticsearch brew tap elastic/tap brew install elastic/tap/elasticsearch-full ``` ### Install on Linux ``` sudo apt-get update sudo apt-get install -y apt-transport-https curl -fsSL https://packages.elastic.co/GPG-KEY-elasticsearch | sudo apt-key add - curl -s https://packages.elastic.co/apt-key.gpg | sudo apt-key add - echo "deb https://packages.elastic.co/elasticsearch/apt stable main" | sudo tee /etc/apt/sources.list.d/elastic-apt.list sudo apt-get update sudo apt-get install -y elasticsearch ``` ### Install on Windows ``` # Install Elasticsearch choco install elasticsearch ``` ### Start Elasticsearch ``` sudo systemctl start elasticsearch sudo systemctl enable elasticsearch ``` ### Configure Elasticsearch ``` sudo nano /etc/elasticsearch/elasticsearch.yml ``` Add the following to the file: ``` network.host: 0.0.0.0 cluster.name: my-cluster node.name: my-node ``` Save and exit the file, then restart Elasticsearch: ``` sudo systemctl restart elasticsearch ``` ### Create Indices ``` curl -XPUT "http://localhost:9200" -H 'Content-Type: application/json' -d

### Overview: This repository implements the sparse attention mechanism from the paper "Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention". It provides an efficient training and inference implementation based on Triton, focusing on variable-length attention and KV cache management. ### Key features: - Efficient training and inference for Native Sparse Attention. - Supports both training and inference. - Includes a toy model `model.ToyNSALlama` for training and inference. - Provides detailed usage instructions in the `ops` module documentation. ### Tech stack: Python ### Why it matters: This implementation is crucial for accelerating sparse attention operations in hardware-aware models, optimizing memory usage, and enabling faster training and inference times.

### Overview: This repository is a curated list of papers focusing on accelerating Large Language Models (LLMs), with a particular emphasis on inference acceleration. The list includes surveys and technical insights into various optimization techniques and approaches. The primary goal is to provide a comprehensive overview of the current state of research in LLM inference and serve, making it valuable for developers and researchers alike. ### Key Features: - **Hardware and Software Co-Design**: A survey on techniques for optimizing transformer inference. - **Model Compression**: Techniques for compressing large language models. - **Performance Analysis**: Insights into the performance of large language models, including optimization techniques like quantization, pruning, continuous batching, and virtual memory. - **Efficient Serving**: Surveys on recent advances in LLM inference serving, including optimization techniques and system insights. ### Tech Stack: The repository uses Markdown and GitHub for its organization and documentation. ### Why It Matters: This list serves as a valuable resource for developers and researchers looking to understand and implement techniques for accelerating LLMs. It provides a comprehensive overview of the current state of research and can be a starting point for further exploration and contributions to the field.

Overview: Qwen2-Boundless is a fine-tuned language model designed for generating responses to a wide range of questions, including sensitive topics, with a focus on cybersecurity and ethical considerations. Key features: - Fine-tuned from Qwen2-1.5B-Instruct - Capable of handling both conventional and controversial questions - Uses specialized datasets including violent, explicit, illegal, and unethical content - Optimized for Chinese language - Fine-tuned using the LLaMA-Factory project Tech stack: Python Why it matters: This model is designed to handle sensitive topics and is optimized for Chinese, making it suitable for applications requiring ethical and sensitive content management. Repository: ystemsrx/Qwen2-Boundless Primary language: Python GitHub stars: 285 GitHub description: A fine-tuned model from Qwen2-1.5B-Instruct, capable of handling sensitive topics. / 从 Qwen2-1.5B-Instruct 微调，能处理各类敏感话题 README excerpt: [中文](README.zh.md)

Overview: NVIDIA NIM Anywhere is a comprehensive toolset designed to accelerate the integration of NVIDIA NIMs into RAG architectures, making it easier to build and deploy effective retrieval-augmented generative AI models for enterprises. Key features: - Native scaling capabilities for full-sized labs and production environments. - Integration with NVIDIA NIM micro-services for RAG pipeline setup. - Easy addition of NIMs as needed for enhanced model accuracy and reliability. Tech stack: Python 3.10 | 3.11 | 3.12 Why it matters: NVIDIA NIM Anywhere simplifies the process of integrating NVIDIA NIMs into RAG architectures, making it accessible for enterprises to leverage their internal data for generative AI applications. This toolset ensures that models can efficiently retrieve relevant external information without modifying the model itself, enhancing the accuracy and reliability of generative AI applications. Repository: NVIDIA/nim-anywhere GitHub stars: 284 GitHub description: Accelerate your Gen AI with NVIDIA NIM and NVIDIA AI Workbench

Overview: This repository contains multilingual translations of "Foundations of Large Language Models" and "NLPBook", including Chinese, Japanese, French, German, Italian, and Portuguese versions. Key features: - Multilingual translations of "Foundations of Large Language Models" and "NLPBook" - LaTeXTrans tool for automatic translation - Original English versions and Chinese translations Tech stack: LaTeXTrans Why it matters: This project provides a comprehensive resource for understanding large language models and modern NLP technology, translated into multiple languages, making it accessible to a global audience.

### Overview: LAMP is a **few-shot-based** method for text-to-video generation. It requires only **8~16 videos** and **1 GPU (> 15 GB VRAM)** for training. Once trained, it can generate videos with learned motion patterns using minimal data. ### Key Features: - **Few-Shot-Based**: LAMP leverages a few labeled videos for training to generate videos with learned motion patterns. - **Minimal Data**: Requires only **8~16 videos** and **1 GPU (> 15 GB VRAM)** for training. - **Training Efficiency**: Achieves high-quality video generation with minimal data, making it suitable for applications where data is limited. ### Tech Stack: - **Python 3.8** - **PyTorch 1.12.1** - **CUDA 11.3** - **Torchvision 0.13.1** - **XFormers 0.0.13** ### Why It Matters: LAMP offers a practical solution for generating videos with learned motion patterns, making it valuable for applications where data is scarce or expensive to collect. The method's efficiency and effectiveness make it a significant contribution to the field of text-to-video generation.

Overview: This project automates the setup, fine-tuning, and evaluation of language models, focusing on zero-friction fine-tuning skills for Claude Code, Gemini CLI, and any ACP agent. Key features: - Fine-tuning of a summarization model on customer support Q&As - Running on NVIDIA GPUs via Unsloth, Apple Silicon via MLX-tune, and free cloud GPUs via colab-mcp - Part of the Gaslamp AI platform Tech stack: Python 3.10+ Why it matters: Simplifies the process of fine-tuning language models, making it accessible and efficient for developers and researchers. Repository: TYH-labs/unsloth-buddy GitHub stars: 283 GitHub description: Zero-friction LLM fine-tuning skill for Claude Code, Gemini CLI & any ACP agent.

Overview: This Python module evaluates Automatic Speech Recognition (ASR) hypotheses, providing metrics such as word error rate (WER) and word recognition rate. Key features: - Evaluates ASR hypotheses - Calculates WER and word recognition rate - Dependent on the editdistance project for computing edit distances between sequences Tech stack: Python Why it matters: This module is useful for evaluating the quality of ASR results, providing metrics that can help in the development and improvement of ASR systems. Repository: belambert/asr-evaluation GitHub stars: 284

Overview: This project is an admin UI for the Chroma embedding database, built with Next.js, providing a user-friendly interface for managing and accessing the database. Key features: - Admin UI for Chroma embedding database - Authentication support - Local development and Docker run options Tech stack: TypeScript Why it matters: This admin UI simplifies the management of Chroma embedding database, making it easier for developers and administrators to interact with the database effectively.