AI Agents & Automation

GPTFuzz is an open-source tool designed for red teaming large language models (LLMs) by automatically generating jailbreak prompts. This process helps identify vulnerabilities and weaknesses in AI models, ultimately enhancing their robustness and security. The repository provides the official codebase for "GPTFUZZER: Red Teaming Large Language Models with Auto-Generated Jailbreak Prompts." It includes datasets for harmful questions and human-written templates, along with a finetuned RoBERTa-large model for judgment. Researchers can use GPTFuzz to generate their own adversarial templates and contribute to building a general black-box fuzzing framework for LLMs.

GraphNeuralNetwork is an open-source project offering implementations and experimental setups for popular graph neural network models. It currently supports Graph Convolutional Networks (GCN), GraphSAGE, and Graph Attention Networks (GAT). This tool is designed for researchers and developers who are working on graph-based machine learning tasks and need to experiment with or apply these advanced neural network architectures. The repository provides clear instructions on how to set up the environment and run examples for each model, facilitating quick adoption and experimentation. It serves as a valuable resource for understanding and utilizing the core concepts of graph neural networks.

gretel-synthetics is an open-source library designed for generating synthetic data, catering to both structured and unstructured text. It incorporates differentially private learning, ensuring data privacy during the synthesis process. The library provides various models, including Timeseries DGAN for time-series data and ACTGAN (Anyway CTGAN) for tabular data, which offers improved memory usage and automated column detection. Users can install the library via pip and integrate it with dependencies like PyTorch and SDV, depending on the chosen model. It is particularly useful for data scientists and machine learning engineers who require synthetic datasets for development, testing, or privacy-preserving data sharing.

gpt-assistant is an open-source experimental tool designed to provide an autonomous GPT agent with browser access, enabling it to accomplish various tasks. Built using Qwik for the frontend and Puppeteer for browser automation, this tool allows users to input prompts, and the agent then navigates and interacts with web pages to fulfill those instructions. It demonstrates the potential for AI agents to perform complex, multi-step tasks within a web environment, such as gathering information, filling out forms, or interacting with web applications. The project is in its early stages, with ongoing development to refine its capabilities and improve documentation.

Gymnasium-Robotics offers a comprehensive suite of robotics simulation environments designed for reinforcement learning research and development. Built upon the Gymnasium API and leveraging the MuJoCo physics engine, it provides a standardized platform for training AI agents. The library includes diverse environments such as Fetch robot manipulation tasks, Shadow Dexterous Hand object manipulation, and multi-agent factorizations with MaMuJoCo. It also features D4RL environments, including maze navigation and Adroit Arm tasks. A key differentiator is its extension of the core Gymnasium API with the GoalEnv class, enabling advanced functionalities like Hindsight Experience Replay (HER) through re-computable reward, terminated, and truncated signals.

GyoiThon is an open-source penetration testing tool that leverages machine learning to automate intelligence gathering and vulnerability assessment for web servers. It can remotely access target web servers to identify products like CMS, web server software, frameworks, and programming languages. The tool offers various intelligence gathering engines, including web crawling, Google Custom Search API, Censys, and exploration of default contents. By analyzing gathered information using string pattern matching and machine learning, GyoiThon can pinpoint product/version details, CVE numbers, unnecessary HTML comments, debug messages, and login pages. It can also execute exploit modules via Metasploit to assess real vulnerabilities, making it a comprehensive solution for cybersecurity professionals.

snake-ga is an AI agent designed to learn how to play the classic Snake game from scratch using Deep Reinforcement Learning. The project leverages Deep Q-Learning, where the system receives state parameters and rewards based on its actions, gradually developing a strategy to maximize its score without explicit game rules. This approach enables the AI to achieve scores up to 50 points with a solid strategy after only five minutes of training. The tool also supports Bayesian Optimization to fine-tune the parameters of the Deep neural network and other Deep RL aspects. Implemented in Pytorch, it offers a robust platform for experimenting with AI in game environments.

Habitat-Lab is a modular, high-level library designed for end-to-end development in embodied AI. It facilitates the training of AI agents to perform a wide array of tasks in indoor environments, such as navigation, rearrangement, instruction following, and human interaction. The library supports flexible task definitions, allowing users to create novel single and multi-agent tasks. It also enables the configuration and instantiation of diverse embodied agents, including commercial robots and humanoids, with customizable sensors and capabilities. Habitat-Lab provides algorithms for single and multi-agent training through imitation or reinforcement learning, along with tools for benchmarking performance. It also includes a framework for human-in-the-loop interaction, enabling data collection and interaction with trained agents. The library uses Habitat-Sim as its core simulator and offers extensive documentation and baselines.

HPSv2 is a comprehensive benchmark designed for evaluating human preferences in text-to-image synthesis. It features the Human Preference Dataset v2 (HPD v2), a large-scale dataset comprising 798k preference choices across 430k images, and the Human Preference Score v2 (HPS v2), a preference prediction model trained on HPD v2. This tool allows users to compare images generated with the same prompt and provides a fair, stable, and easy-to-use set of evaluation prompts. It supports benchmarking models across various styles like Animation, Concept-art, Painting, and Photo, and offers functionalities for custom model evaluation and preference model assessment.

RynnVLA-002 is an autoregressive action world model developed by Alibaba-Damo Academy, designed to unify action and image understanding and generation within a single framework. It integrates Vision-Language-Action (VLA) models and world models, building upon its predecessor, WorldVLA. Key enhancements in RynnVLA-002 include a continuous Action Transformer, wrist camera input and generation capabilities, and state input. The model has demonstrated high performance, achieving a 97.4% success rate on the LIBERO benchmark. It offers both VLA model capabilities for generating actions from text instructions and image observations, and World Model capabilities for predicting future frames based on current frames and actions. The project provides models, training code, and evaluation code for both LIBERO simulation and real-world LeRobot experiments.

Sagittarius is an innovative open-source tool designed for exploring the voice and video capabilities of GPT-4 and Gemini models. It provides an online platform where users can interact with these advanced AI models using both voice and video inputs, offering a real-time exploration of multimodal AI. The tool is accessible directly through a web browser, eliminating the need for any installation. Users simply require an API key from either OpenAI (with access to the gpt-4-vision-preview model) or Gemini to get started. Sagittarius is noted for its speed and support for multiple voices, making it a versatile option for developers and enthusiasts interested in cutting-edge AI interactions.

IndoNLU is a comprehensive collection of Natural Language Understanding (NLU) resources specifically designed for Bahasa Indonesia. It features 12 distinct downstream tasks, offering a robust benchmark for evaluating Indonesian language processing models. The project provides code to reproduce results and includes large pre-trained IndoBERT and IndoBERT-lite models, which were trained on an extensive 4-billion-word corpus (Indo4B) comprising over 20 GB of text data. Developed through a collaboration between universities and industry partners, IndoNLU also offers access to the Indo4B dataset and various FastText models. It serves as a vital resource for researchers and developers working on Indonesian NLP.

InternAgent 1.5 is a sophisticated autonomous system designed for end-to-end scientific discovery, encompassing both Algorithm Discovery and Empirical Discovery. Building upon InternAgent 1.0, it structures scientific inquiry into three interconnected subsystems: Generation (hypothesis construction via deep research), Verification (methodological evaluation through solution refinement), and Evolution (evidence-driven refinement using long-horizon memory). This framework achieves leading performance on scientific reasoning benchmarks like GAIA, HLE, GPQA, and FrontierScience, demonstrating sustained autonomous optimization over extended discovery cycles. It supports diverse workflows, from agent memory and reinforcement learning to dry-lab simulations and wet-lab experimentation, across Physical, Biological, Earth, and Life Sciences.

Stephanie is an open-source platform designed for building voice-controlled applications and automating daily tasks, mimicking the functionality of a virtual assistant. It provides a flexible framework for developers to create and customize their own voice-controlled systems. The platform emphasizes its open-source nature, allowing for community contributions and extensive modification. Key features include voice control, task automation, and an intent prediction algorithm called Sounder. It supports Python and offers detailed documentation for installation, configuration, and usage, making it suitable for technical users looking to implement custom voice solutions.

LibKGE is a PyTorch-based library designed for the efficient training, evaluation, and hyperparameter optimization of knowledge graph embeddings (KGE). Its primary goal is to foster reproducible research and facilitate meaningful comparisons between KGE models and training methods. The library is highly configurable, easy to use, and extensible, providing clean implementations of various training strategies, hyperparameter optimization techniques, and evaluation metrics. It supports common KGE models like RESCAL, TransE, DistMult, ComplEx, and ConvE, with explicit exposure of all parameters via well-documented configuration files. LibKGE also includes GraSH for efficient multi-fidelity hyperparameter optimization on large-scale KGE models.

Keepsake is a Python library designed for version control in machine learning, enabling developers to track and manage their experiments efficiently. It automatically uploads files and metadata, including hyperparameters, training data, weights, metrics, and Python dependencies, to Amazon S3 or Google Cloud Storage. Users can retrieve this data via a command-line interface or within a notebook environment, facilitating experiment analysis and replication. Keepsake supports various ML frameworks like Tensorflow, PyTorch, and scikit-learn, storing data as plain files for easy integration into production systems. It also offers features for comparing experiments, committing code to Git after the fact, and loading models for production use.

Grizzly AI leverages AI agents to significantly reduce the time and effort required to respond to Requests for Proposals (RFPs), bids, and tenders. The platform helps teams quickly understand the scope, scoring criteria, and requirements of a tender by mapping the tender pack. It then drafts responses by matching existing content (past bids, case studies) to the new questions, ensuring relevance and consistency. Before submission, Grizzly AI scores answers against evaluation criteria, flags missing evidence, and identifies areas for improvement. Post-submission, it integrates feedback to continuously refine the content library, making each subsequent bid more efficient and effective. This ensures senior staff can focus on won projects, allows for bidding on more opportunities, and ultimately increases win rates by aligning responses with evaluator expectations.

lexvec is an open-source implementation of the LexVec word embedding model, designed to achieve state-of-the-art results in various Natural Language Processing (NLP) tasks. It functions similarly to other popular word embedding models like word2vec and GloVe. The tool offers pre-trained vectors derived from Common Crawl data, available in both subword and standard LexVec formats, with options for cased and lowercased words. Users can compute vectors for out-of-vocabulary (OOV) words using a binary model. LexVec supports both in-memory and external memory training, making it adaptable for large corpora. It also incorporates subword information through character n-grams, enhancing its performance. The project includes evaluation metrics against other models like fastText, word2vec Skip-gram, and GloVe, demonstrating its competitive accuracy.

llm_benchmarks is an open-source repository offering a comprehensive collection of benchmarks and datasets specifically designed for evaluating Large Language Models (LLMs). It covers a wide array of assessment areas, including knowledge and language understanding with benchmarks like MMLU, GLUE, and Natural Questions. The repository also includes resources for evaluating reasoning capabilities through datasets such as GSM8K, DROP, and AGIEval. Furthermore, it addresses grounding, abstractive summarization, and content moderation with benchmarks like ACI-BENCH, MS-MARCO, and ToxiGen. This tool is invaluable for researchers and developers looking to rigorously test and compare the performance of different LLMs across diverse linguistic and cognitive tasks.

MADDPG offers a PyTorch implementation of the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm, specifically tailored for multi-agent reinforcement learning in environments with both cooperative and competitive elements. This tool corresponds to the research paper "Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments." It includes a quick start guide for running simulations, such as testing the algorithm on a 'simple_tag' scenario with pretrained models. The implementation allows for training agents, like predators catching prey, within the Multi-Agent Particle Environment (MPE). Users can also modify environment settings, such as switching between sparse and dense reward structures, to suit their experimental needs.

Lobe Vidol is an open-source platform designed to make virtual idol creation accessible to everyone. It boasts an exquisite UI and integrates support for MMD dance content, allowing users to bring their virtual idols to life with dynamic performances. The platform also enables seamless conversations with characters through text and video chat modes, offering an immersive interactive experience. Users can create custom virtual idols, set touch responses, and upload VRM models. Lobe Vidol supports a diverse range of multi-model providers, including AWS Bedrock, Google AI, Anthropic, and more, ensuring rich and varied conversation options. It also features a character and dance marketplace, TTS & STT voice conversations, and is available as a Progressive Web Application (PWA) for a seamless experience across devices.

LobeHub is an ultimate space for work and life, designed to facilitate the discovery, creation, and collaboration with AI agent teammates. It takes agent harnessing to the next level by enabling multi-agent collaboration, effortless agent team design, and introducing agents as the unit of work interaction. The platform offers features like an Agent Builder for personalized AI teams, access to over 10,000 skills and MCP-compatible plugins, and Agent Groups for scalable collaboration. LobeHub also emphasizes co-evolution through Personal Memory and Continual Learning, allowing agents to adapt and grow with user workflows. Additional features include a desktop app, smart internet search, Chain of Thought visualization, branching conversations, and support for Claude Artifacts, file uploads, and multiple model service providers.

ma-gym is an open-source collection of multi-agent environments built upon the OpenAI gym framework. It is specifically designed to facilitate research and development in the field of multi-agent reinforcement learning. The tool offers a variety of pre-configured environments, including Checkers, Combat, PredatorPrey, Pong Duel, Switch, Lumberjacks, and TrafficJunction, allowing developers and researchers to simulate and test multi-agent systems. It also provides a multi-agent wrapper for existing OpenAI environments like CartPole-v0, making them accessible for multi-agent experimentation. The project emphasizes ease of installation and usage, with clear instructions for setting up the environments and running simulations.

Diffuman4D is an open-source project that provides a framework for 4D consistent human view synthesis from sparse-view videos, utilizing spatio-temporal diffusion models. Developed by zju3dv, this tool allows for high-fidelity free-viewpoint rendering of human performances. It includes scripts for inference, data preprocessing, and reconstruction of 3DGS and 4DGS models. The project also offers a meticulously processed DNA-Rendering dataset with re-annotated labels, including foreground masks, 2D/3D skeletons, and camera parameters, to facilitate further research in human-centric 3D/4D generation. An interactive demo is available for users to experience immersive 4DGS rendering.